<%BANNER%>

Diffeomorphic Point Matching with Applications in Medical Image Analysis

xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID E20101123_AAAACJ INGEST_TIME 2010-11-23T11:55:19Z PACKAGE UFE0011645_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
FILE SIZE 2176351 DFID F20101123_AABFTE ORIGIN DEPOSITOR PATH guo_h.pdf GLOBAL false PRESERVATION BIT MESSAGE_DIGEST ALGORITHM MD5
cbfa8b66d8f0965f6ef0372aa53d262c
SHA-1
b655f51c8da656dd767f2bae59fe11da2cab6d47
25271604 F20101123_AABFSQ guo_h_Page_102.tif
f14b2bb9a0834ee01e0f8b7557b94482
4da9e95054cb9d327c274d041ed18ff878fb35f3
814 F20101123_AABFTF guo_h_Page_078.txt
33823d0aefe7d54a05a813e700cbf604
d82c52981cb08ecd25d9e1249f1e03aa00ad7126
96878 F20101123_AABFSR guo_h_Page_082.jpg
c45714503c764ef4781e9e6e609b5669
2662773587c10bee048c5a2626a1762d65bbc127
29647 F20101123_AABFTG guo_h_Page_077.jp2
f577bdb95d8aa7e5800a42731a8573f0
7184020dabe76b22dc7c592f6f98c204ce15895b
F20101123_AABFSS guo_h_Page_044.tif
efe29c816a7125b2aa0ed88d5f29ef8b
e45c1665448c1c92b004a8a2ed8801d0939cfe37
24318 F20101123_AABFTH guo_h_Page_045.QC.jpg
bc00eaab409199fc58de7da12b3cbbfc
4af87dd172c039cf45a9d746a5963c9af73395a4
61897 F20101123_AABFTI guo_h_Page_091.jpg
6ce7e18616ae60a684538c794f4c9ff6
9dbcd715c064a6fdee096f984424afaf4e460b1c
6171 F20101123_AABFST guo_h_Page_039thm.jpg
75d56a14bc16e1a0521687e24ea62bf2
8a8c52aa87a6ab79a9e75befd393e24abdb4e1c1
5981 F20101123_AABFTJ guo_h_Page_019thm.jpg
f8c5902fc4b118a44d6dfa3c4c667761
6e56be137c50cd73c408708e0d5ba0ec4137fbe9
3131 F20101123_AABFSU guo_h_Page_050thm.jpg
64d9fc5a69880e7a51054b5df77e1ae6
bbaa2d17e2a6dc0f5720a01e02efc4cd664b0f4b
1053954 F20101123_AABFTK guo_h_Page_076.tif
7bb043ecab5310aeeb9328717b54adb3
aaaa7ed30708b011ad02869b267eaceb5fa6bd42
16666 F20101123_AABFSV guo_h_Page_044.QC.jpg
0f283d583621f07bb805702b8488df48
ec2a74d6a76118420bd813eeb55a307f421caffa
10224 F20101123_AABFTL guo_h_Page_002.jpg
454f258f3def5500bf61323c71ea897e
a3446b0ab804f2583254d6aea5f72a1da14fda0b
1178 F20101123_AABFSW guo_h_Page_069.txt
2934469d19cc4a947c441ac0d587cced
998f7ee80f76027a310db3cc544c136e948ae952
75290 F20101123_AABFUA guo_h_Page_012.jpg
6072e4627386a32eac3fb5ff981d998e
640a73d2d7835dcbdd468a0d3bba5431e1dff12f
210 F20101123_AABFTM guo_h_Page_098.txt
6841c17dbf5ed4e07de78275923673ab
4c49c2538a0bdd784e30d276a9bf60d28761d9bf
1051982 F20101123_AABFSX guo_h_Page_112.jp2
62ece9d0afbf7bb29cfa6df7e2d96495
93e1e115a61005c0d8858af351c1a9f073ff29a7
83857 F20101123_AABFUB guo_h_Page_013.jpg
0a0acf147bdf04d0abfca4c2af5d7c0e
8764d1f1b47a22f4411ac146bd14aa193b1c4c9c
3650 F20101123_AABFTN guo_h_Page_005thm.jpg
754ae5559b44441c9a8bd49f695de873
e8c15f8b2f912ed09c41e635aa5b9d86108da75a
F20101123_AABFSY guo_h_Page_077.tif
828ac87839fd6fa017f7bfd60b426b54
ccc808517a08bbaa2607b7af9bc513d1dee0f55b
84307 F20101123_AABFUC guo_h_Page_014.jpg
e3a5fb88a61c49933c541b63a7f9c586
83ae3661678f84fc5b51b6caf147717bb84af56b
4323 F20101123_AABFTO guo_h_Page_033thm.jpg
4eef79ea6b18d07430c3ca7f2d21b651
8b71997dcf539b7bbd9237da35cf3fc3fc8b7d20
2218 F20101123_AABFSZ guo_h_Page_070.txt
c0054e7c2a85ca88f9109ed1f48ac3c3
45e50e186fd0643034fc0555ededbd3f83a9ffc2
58841 F20101123_AABFUD guo_h_Page_015.jpg
983ec20f2fdb2f48fd9d8b97cc4c35b7
7133935d5b0ad983f0eab8fc87139f09e5973107
173256 F20101123_AABFTP UFE0011645_00001.xml FULL
04a308dc77ee85b457f0c231ae42163e
1f00b9d4299fdbcffc08f57b7d5ba99f7d321f24
71682 F20101123_AABFUE guo_h_Page_017.jpg
0f47ea50ee7ae204ff3c1cb62baa2a34
b493dd98aa66350635076dda7c56959c470d996d
67827 F20101123_AABFUF guo_h_Page_018.jpg
660ef4992dc488b3a268da704ad26d67
316c54fa6e2234805d24e8e27730b5568fc1d3b7
64672 F20101123_AABFUG guo_h_Page_019.jpg
a78ccef6b2780eb4f8014cbd182582ae
d4f1c677674660c78a8b78d87094a62b0d69488f
23914 F20101123_AABFTS guo_h_Page_001.jpg
1936f85e9524f79f293dd5f863d7796e
043c6d5dfbd0d0b613d73b374baf316146dd42ac
864903 F20101123_AABGAA guo_h_Page_092.jp2
6c1f9ce932f42d26dd86a72473c42b7f
355486d0a4f01880dd73d77367914a9fa4fd00ac
29255 F20101123_AABFUH guo_h_Page_020.jpg
3d3205434d5bb5f93deb910ed12f6d16
ee7f8759112204bf3f37040a68596af2cb9461ff
68906 F20101123_AABFTT guo_h_Page_004.jpg
77a56dd41489be85beaa2df9fdf4369e
bee86e1ea597c82bfa43dfecea4b99a492d68533
786393 F20101123_AABGAB guo_h_Page_095.jp2
4f562b0fbe70b93e388c5c34a0991c5a
f51a36d3e4805bad933f366c486e58a8cb97986b
51239 F20101123_AABFUI guo_h_Page_021.jpg
c447d94675c8bef92f99dac0694ae777
739fcd477bd5d4bb8d118c94f7f1d7e51fc7b7e8
985399 F20101123_AABGAC guo_h_Page_097.jp2
6cd2ada6b61f9406f252c76ac413a78c
f906b672fce363794e9999fdc125c9f0117ffb20
82806 F20101123_AABFUJ guo_h_Page_022.jpg
85039c396f127a3ccca9449fad48b3ed
f57ad013ea52d06916483ab8a5cbb197016c6aa5
73220 F20101123_AABFTU guo_h_Page_006.jpg
3be26dd425185c3a213d7b01e1c7c709
4c5cedaa00b94896856860d678b3c1b29763eb23
1051970 F20101123_AABGAD guo_h_Page_100.jp2
79db810cdcc3d58082b06b8105086d80
5a1772300a34864ed3799a75e9106553fa0c6e60
83977 F20101123_AABFUK guo_h_Page_023.jpg
62e4aa11309abd164bd0644638f0891c
e40dfa89302bede770c08126b33bc34ae043a0bb
42622 F20101123_AABFTV guo_h_Page_007.jpg
3d5daa013c19f9a5fa3440ec6624ba35
3033162472f7fd6df542613c45c08e56aee51675
469313 F20101123_AABGAE guo_h_Page_101.jp2
82c3ea9dc2a6191dd909e6f978e69d33
d92c40398f3b02f7ae4ac6bca6beb22082175dd5
35620 F20101123_AABFUL guo_h_Page_024.jpg
fa04a14aa65b37f4597d8b30530b556e
4bf1d83f2eb4a290bc284ba10c909b96dc31be47
28482 F20101123_AABFTW guo_h_Page_008.jpg
0fb3cb22bcfe21c7306c8a7bd0a73671
78ee911f173a3330df402aa32fd0b41ddfeafc24
400803 F20101123_AABGAF guo_h_Page_102.jp2
a459fa8918d699e24a4949b31a56cda7
3eb4c98f73bd2c53e363a9792f56da053eb68428
82273 F20101123_AABFUM guo_h_Page_025.jpg
2aa1db956103c1bd9ab480020dab045a
f6c674bddef5e31af03e4b72e83f68866c03ddeb
63604 F20101123_AABFTX guo_h_Page_009.jpg
a45d09a7f921f85f6db4529a97b80402
596a110a8e56a267648a8e50f20c2280753c7977
901784 F20101123_AABGAG guo_h_Page_103.jp2
6cdcfc08d01d61daa585b860e7fa2df4
9e2b6dcedcbf08dfc9aa8ee5d16e9c5ea168c9b0
66539 F20101123_AABFVA guo_h_Page_042.jpg
250289c9fdf311fe5bcce087ced952d9
4f071b7dfa47c3b65e2130310f6fa1467f1c7898
75650 F20101123_AABFUN guo_h_Page_028.jpg
0ac25fedf06995e86a7f672a55997c12
90ac0d6707b0e3ea33a4a7f21847c6fc6d19b8fd
59528 F20101123_AABFTY guo_h_Page_010.jpg
bb27f4caf7cfe68ba1490d6380f9e7a6
71fa975c806c0740be94dc65bc064a12173cde75
469289 F20101123_AABGAH guo_h_Page_105.jp2
85b6798d129d3d1b908f6936d6ba5fa2
2f05899abe8690932c0ced43e91509751b2bc561
78980 F20101123_AABFVB guo_h_Page_043.jpg
589343c500bb3c3a02e5f363daa3d881
93829eb919fd8fbe0150f1a01c3578917f1f6bbc
69923 F20101123_AABFTZ guo_h_Page_011.jpg
7d8a0f0a9968baa029015462e2bad331
af951ec6075f879b28733acf5d60b9f25fd33db6
552661 F20101123_AABGAI guo_h_Page_106.jp2
fd05180151315c0ba46f9a3663f11c1a
6f6f5898cc5e7bc3b4cae443e24a37a01b3845ad
54488 F20101123_AABFVC guo_h_Page_044.jpg
e7811be3fd622fd2638aa737085146b7
95061cc21d65b1aa482471895ddd6a53d79278e4
78532 F20101123_AABFUO guo_h_Page_029.jpg
63e7c4a809151cc03cecc30309546eb1
dd054d6c6babd720f8fc5e6dd29279a596195a02
389318 F20101123_AABGAJ guo_h_Page_107.jp2
f0e53fac88436c94fe1774090fd59ae5
1a3a20f50a8c314c5060b4f105b5de53fe099d42
77111 F20101123_AABFVD guo_h_Page_045.jpg
0916d34d39a80647b638bccf4571149d
09c446e6a55ef105b146a8ff1aae48066ff203c7
68890 F20101123_AABFUP guo_h_Page_030.jpg
9e3009ebaa765f3f4e69fb960542b61c
46b1e4af7bfa5e574ab32864b2721cf7a20556d3
294034 F20101123_AABGAK guo_h_Page_108.jp2
ebd1590f9fec979fd65f28bb68b5c12f
10b149140b673fc8aba4f23ef7d41a5f7d1f3d90
47916 F20101123_AABFVE guo_h_Page_046.jpg
79d68b870f96c2f4e85e9922bc7d7ec6
a3d14e21a0cba9176a0e3d8ff7bc604dd1a02be2
40901 F20101123_AABFUQ guo_h_Page_031.jpg
aa33ae648273daa1ea0cc37c4e66d9f7
cef907ff1b7b318530238f7e3d978b1f9634795c
F20101123_AABGBA guo_h_Page_011.tif
23a4918aaa63e152fe9bbb090ade66d6
08e35a354c661aaedb2d483b7a629e8a85c806c1
1046087 F20101123_AABGAL guo_h_Page_109.jp2
cb6b847c6386541ce6253e4b20b5ad84
d1cbb721b5ae26ca9feabff07a0e9d226d1701eb
62552 F20101123_AABFVF guo_h_Page_047.jpg
5c7bd1bedf8fb1833206d145ec83bb20
15cf24f0e6ededa62cde9497fe00f35d5115e1f6
60811 F20101123_AABFUR guo_h_Page_032.jpg
a825dec2444b78ba484033515cb910f6
c58e4a60448533ddb991c12e1dfa5feb40124a8e
657239 F20101123_AABGAM guo_h_Page_110.jp2
907ef5eb766119f5b5e746c1efec3e0e
0f55568b22804f4935cacd78c60c5cdcda03fa9d
43217 F20101123_AABFVG guo_h_Page_048.jpg
ca88434e22714effbf9a86d76500461a
6e333dc295d95fc3d7efd083eb25e719921319f3
43671 F20101123_AABFUS guo_h_Page_033.jpg
b220c93b850b6c6289442f523b3c1e04
6cfcfe6cd54c4a447d983a71cbc21d72061cfce1
F20101123_AABGBB guo_h_Page_013.tif
29af65430222136b20693af96c73ec04
4bace4fe913d9cb7f9e890a34d0d42d7263558a4
1051856 F20101123_AABGAN guo_h_Page_111.jp2
d98756fd11bb1479ceeab82e853a5dc4
ed5644c4086af45f07947076d92a54804b78ab19
40720 F20101123_AABFVH guo_h_Page_049.jpg
0ded08eaabf3a18e866bb6340c17053b
557f572c34daf7e8066fe97c560887c379ce34f2
41371 F20101123_AABFUT guo_h_Page_034.jpg
fe0fd9c7f98936ace0c4076078e4737c
73ceeb85b7522731c9eb292e225c60b880807294
F20101123_AABGBC guo_h_Page_015.tif
0b2b2fc996ceabbe04392b5c43527829
d0bc0b671ccebe2002fa44734ff6ba0c8300492d
1051934 F20101123_AABGAO guo_h_Page_113.jp2
79d48cb6f10540ffd0b3278afd202714
7da87eaf40879b52fca1b60c77de1d22848adbca
40392 F20101123_AABFVI guo_h_Page_050.jpg
003036aafd134f1e920544a6ef629fa0
fb4cf1a822744963f21372b75b7c4c9a59b9e104
60626 F20101123_AABFUU guo_h_Page_035.jpg
0fb1bb9c0d3924b6e6f8bef30635a15f
e6be24254f6da3898911c3a539e58b5371efd10f
F20101123_AABGBD guo_h_Page_017.tif
bad936ab10c44402cababc49a7babc23
19c71b06bb0427bbab10a03201e2815ffa045941
1051962 F20101123_AABGAP guo_h_Page_114.jp2
5ec1ee3481d25394b3f3709ed53ceb0d
1836b81acc0c6efc395af26b6efd81d08c0d0a2d
33466 F20101123_AABFVJ guo_h_Page_051.jpg
b36779b53eb16b42ddc938c24f49f342
7bcf1e291b3bf00ce419cf71a146c5532006863b
F20101123_AABGBE guo_h_Page_018.tif
b7026f99a43277c46b2d66866ea5041a
ea5edbce9f541ce4ce4af71cddbeb0e01c7058bc
1051955 F20101123_AABGAQ guo_h_Page_115.jp2
26c2ca25310d892aa5dffe43242f249c
9ee450315034d5673812fc464a6661fa2fe0a47a
69992 F20101123_AABFVK guo_h_Page_052.jpg
bd963333dda7c8f91f6001ec8cc4ba52
b7699e0541f914b621a6cc92f7e8ea6cec225603
71382 F20101123_AABFUV guo_h_Page_036.jpg
b7a45ea16247f007b6f9a9adbb860a7e
502cfa48b2634485de67fe076325df5ae4c18f83
F20101123_AABGBF guo_h_Page_019.tif
5c42ef7c01c8b87754fc07f0e6516e63
6cf290a1ee34bbc90addf8221e3a1116b329cd0d
237302 F20101123_AABGAR guo_h_Page_117.jp2
02f9c9ad8dc552116d529ccd86f7de98
a75e6c4cf7d068309d8debb451a65f8b9db793f1
69199 F20101123_AABFVL guo_h_Page_053.jpg
052dcda30c5f3f388c482d019646e020
88366c916d3b58989b1776cb9a9bd9a17b07da3e
36949 F20101123_AABFUW guo_h_Page_037.jpg
ec47d0eaebea117bf6567a6fcf112ae5
90b4e6209f8b85799b8f6e855ae17efff83119d2
8423998 F20101123_AABGBG guo_h_Page_020.tif
b12e9c5e01f2dc2d5b7e9d607c72fd38
57667790120a9f6926b8d5391e7fa036f554280c
78344 F20101123_AABFWA guo_h_Page_072.jpg
1f744e1f1bcbdf531482e89ca5eb71cf
41039964659cc5478c2c392b33004fb70718420c
F20101123_AABGAS guo_h_Page_001.tif
c9a742cd774c223669d304e2e574bfb2
fc2f9b1c06863a39b31002f94cd19666f2afdc05
64373 F20101123_AABFVM guo_h_Page_054.jpg
6c04fb9499b2e3af1197d999fd953010
b61f6eed7cc03e7cb260228a82bbe8281166b1ca
69650 F20101123_AABFUX guo_h_Page_039.jpg
863585caf86128b9bf4d722ff8fe9eb8
6234d23fdd7b695f2334c1a7169328f3c47c78dc
F20101123_AABGBH guo_h_Page_022.tif
8f3b8d58a1d4fe308cc1f7e117892af0
53a36c360cf22c8e2c8c84c8b24a2f1896311ed1
75603 F20101123_AABFWB guo_h_Page_073.jpg
cf1024f41f5062c5ce5f7313756c691c
ed8a2ddfbdb4fb89cf26394a694478c59b1bea57
F20101123_AABGAT guo_h_Page_003.tif
3610dc21f70f0469dafcd23a3373e94d
1cc1e7815b498ee2791e51421a31e50725876c2f
49970 F20101123_AABFVN guo_h_Page_055.jpg
b712d9823de0cf7a8c39d367b10f15b7
1bf7b11f7a82b74252b02ec30ce0ede851280d0e
81125 F20101123_AABFUY guo_h_Page_040.jpg
b0587be7a93402a78c552e084f4e00f2
ffb3cf73e5202d227b7e8f2e716fff543850d3a5
F20101123_AABGBI guo_h_Page_024.tif
6b51047a8699e47f8c7995f5e45dbed8
8f0ed09c82b34841f1049d3cd450712b9f35f974
64799 F20101123_AABFWC guo_h_Page_074.jpg
cc6f61d8de78c0450d404d641fd41a38
5fa6b2b767181e74f5a9312d1c32018d845a3976
F20101123_AABGAU guo_h_Page_005.tif
92eabbe776e387f0e49799e2e95179fc
048dcf2f35bd2b4a20a86bce77ff56799bcef0cf
56196 F20101123_AABFVO guo_h_Page_056.jpg
417dfb14f7f8fc843a581e89c258e80d
2cc0cd679f1e2bb749a38ad86dd7aeac0c15a5ec
52900 F20101123_AABFUZ guo_h_Page_041.jpg
c769cf53f36119c3848ea99ea74b0765
002d8306779fe90fde52356290ff5004c049e689
F20101123_AABGBJ guo_h_Page_025.tif
a023ec6383f6f71751e8cff535538c4a
d95c49f4e9298a12b82f32225bfd663b439e1ece
27858 F20101123_AABFWD guo_h_Page_075.jpg
f0e5925972611cae230114df4ce0bf57
68cab882d4421d1a5de9de570126f2c494c4e8b6
F20101123_AABGAV guo_h_Page_006.tif
8af1cd32906d056dc743f81f6545e036
b9674078e33735f86d4f29ef639183d31b7d69de
44007 F20101123_AABFVP guo_h_Page_058.jpg
226b281e1f765665f20bbab8f29a4fcc
9ac5d1b0cdf737d641a7c981fbcabcc8b2f3f34f
F20101123_AABGBK guo_h_Page_026.tif
0b2215066526fb7cba434f3af6068cc0
409f121ace45590ed78f5d111280ca82faf43a34
25061 F20101123_AABFWE guo_h_Page_077.jpg
4993be267d3fdc052cd1f8a12f3b363e
3851fec22053b5027773354c5a53dde7ccbe9c49
F20101123_AABGAW guo_h_Page_007.tif
1cff8daad284e3790d0502533801dc07
db178ef1e1be9a3146f543340d2740e61a935b1c
71527 F20101123_AABFVQ guo_h_Page_059.jpg
0c69d2f09178757a10cfd4389279edb9
6b136df781f2bd85de8728fbe74e61ccf718faf2
22363 F20101123_AABFWF guo_h_Page_078.jpg
7b6878de3dd10c6f2492715a967d38ff
4231c4f405bab6749c447a7a6c458ffb57f8ca85
F20101123_AABGAX guo_h_Page_008.tif
0a8b10fd08402ae6ff01cd496c495e28
100f9431fd26feae007eec7365af69efef9a46c4
70587 F20101123_AABFVR guo_h_Page_060.jpg
4e3fa7fe314e9f294835a82e3e7646e5
cbabf16d47bcba6df0533c034a76b70775df0c15
F20101123_AABGCA guo_h_Page_050.tif
1cd10e9d71da5dd7f393509fe9b66bab
19fdce6425a9ccbeb5ef47cddbf50f10092aeb77
F20101123_AABGBL guo_h_Page_027.tif
ad725c15cab1cb51b4af7ab6d639668f
f5b6fd7c2c2b468c466c8d5d67c276ec3c5eb225
23332 F20101123_AABFWG guo_h_Page_079.jpg
f37901f07d3f095303ac3c54be26ccbd
4667ce125bad7f19eb820e3ef0907b396e7777f5
F20101123_AABGAY guo_h_Page_009.tif
c67ffba96fb9801b4c209ca610060884
1df9853c5fe2f1e6192c6c654674460564dce562
37700 F20101123_AABFVS guo_h_Page_062.jpg
fe98e7f3e23d302ad47154b88948b7f8
91c042b758fe5284b465f78760432926eb7b4ba8
F20101123_AABGCB guo_h_Page_051.tif
880f0b9e289bec9c40ffa1a3b9f92764
026de9826a1021e62ab866e0d49d994781d3b8e6
F20101123_AABGBM guo_h_Page_029.tif
b09747436b5e90a1dd7eed67c22f06e8
1a62ed46d443d6dde7f59f0c9fb82b0ea3892ed2
75763 F20101123_AABFWH guo_h_Page_083.jpg
400cb6c1a5b0c89de4f5e6c4e755c27d
256ffa01d9bed675a8756c2035566917e4d75a78
F20101123_AABGAZ guo_h_Page_010.tif
c0a93fe7b7d2e6f6ae332b22a74dc5f4
0f3b9a46a61eefca629244327113862b4cf6ba96
56895 F20101123_AABFVT guo_h_Page_063.jpg
7f9c4e88856026146cfa7d534a474759
88bc0937e9fc38d963b07c97315ce6dc54c3db73
F20101123_AABGBN guo_h_Page_031.tif
7e5bb6269fa6331b1735458f3f70986d
386f83e5f993c6e46056637884475e1a02dbec3c
57906 F20101123_AABFWI guo_h_Page_085.jpg
a32a0a97367a8e64d593b8e7bb5a6baa
c745c89bbd2cbdb5afa42db8e60fa2c15babf86a
57201 F20101123_AABFVU guo_h_Page_064.jpg
c55a09ba21596d86faefbfd8d3998cd9
bbe31de2a56d5a3cb334420ca1262d8b16c05c7f
F20101123_AABGCC guo_h_Page_052.tif
4fafae71185d4ff76e18127e98456622
bec1f356b9cabb2123524ff0791f519dc616f6be
F20101123_AABGBO guo_h_Page_032.tif
dd570636a3a0c75be04b21f116bedd0d
cb7a185f277c44dcaa17504d51d1df60d6394221
57099 F20101123_AABFWJ guo_h_Page_086.jpg
2d5ef917e9ad5c676d8c8cf5068a7ccd
44b0b9e59474d316ee341c8db5ba257f8ac33f8f
65989 F20101123_AABFVV guo_h_Page_066.jpg
a077be6a934f36e4cb879ca2fedeb99a
3fb390c1969b4a7f604451c1076149878b611bdc
F20101123_AABGCD guo_h_Page_053.tif
44d7ce8a5389c0cf0534acaf42a0b9bd
a6dfed9ea4b008d3237d8c337b13d2bbafb820bd
F20101123_AABGBP guo_h_Page_033.tif
c80e7c8320fa1e8e1a9ed95fb30c6010
466c949feebdcb0d70547094e6592e7014a94496
71713 F20101123_AABFWK guo_h_Page_087.jpg
788e379839678ed41ac9f63a7c4e0a6d
9b17b01b9ffef07b81c8ad677485f467cb14965d
F20101123_AABGCE guo_h_Page_054.tif
946f7b84a3311984db0ee40453148935
2dddac4b97c558e1be4c98bd514eb77b9a693fc0
F20101123_AABGBQ guo_h_Page_035.tif
965d1cf442bedc2e8d9636c7bc55c6a9
061b5b18de2315dcda533366cee578a3a7082104
72538 F20101123_AABFWL guo_h_Page_088.jpg
b6d3acde07c3d90955757f9c53a813d3
fe9c9c41b73a7b33e73c20e911f1b38447866b4a
53479 F20101123_AABFVW guo_h_Page_067.jpg
7ecd3de26f5ceac654baf52eba742fe9
133b612c0bc33169e4e307f8784155626b72fe33
F20101123_AABGCF guo_h_Page_055.tif
abef7cc73d93a37fbc47f29a150135b2
3a6aca1d86956a329a59c80854975ef158495217
F20101123_AABGBR guo_h_Page_036.tif
6e984db276e7d30ef21acb81739deca0
14e2135ce5be6d801fb7f2453de80697d0d737d5
72225 F20101123_AABFWM guo_h_Page_089.jpg
3f20b7289bb1e5b871965ea93fdbe9a2
71123a68378f9b4a32754fb3abc260e84194c5ad
60138 F20101123_AABFVX guo_h_Page_069.jpg
209f388428f0088a0f458c14f1d64cb2
6e1ae982a1c95b1826498c76c0b4514ccf98edae
F20101123_AABGCG guo_h_Page_056.tif
610645ca7aef2a664c1378375e7da196
3a3928fe71c0934de8dcab80639e440327a57c7e
28888 F20101123_AABFXA guo_h_Page_106.jpg
0a7a22c16ed0b5ed10633f25238b636b
78871897a222d44ec65a9a3f1e6061efe0652743
F20101123_AABGBS guo_h_Page_037.tif
2a82865f94533ac0d202d4715be256ba
3d4eeb6760dd0b4f26693fca0f16b30600f8c820
76730 F20101123_AABFWN guo_h_Page_090.jpg
eb3be8f4b2e9c8f88708e6a8231ed27a
0519015a3a37a992cd28e4f601f099b4a96ddaac
67273 F20101123_AABFVY guo_h_Page_070.jpg
1b44cbbc3515a3c782a058bf98a1270c
3858a948da91db73f97ed78025cc08f32e40b309
F20101123_AABGCH guo_h_Page_057.tif
73ce9293f1d1250027989280e7976ed9
c80a396eec05782e61ef6da5a0278b00d3b121fc
24448 F20101123_AABFXB guo_h_Page_107.jpg
4cfb2b2b48dbe1eb5e189b66e45b6cb9
cd3d8df192a28ac8dd606a5ab4ed97f18f026e60
F20101123_AABGBT guo_h_Page_038.tif
c934b2f209e7168e1fe82f61cfb4c65b
330bd850104cd0870863790f9a25f12a17b6937c
28346 F20101123_AABFWO guo_h_Page_092.jpg
97855b69b87f86e908eeb78175b24df7
2ecc565ebed3ebfe27ce8faf6c9bd6fd83a636b3
93121 F20101123_AABFVZ guo_h_Page_071.jpg
ab84997b02870eff6f99a0e9a2e6d915
297732baaca55c6cab223f52280f8e8a76bca184
F20101123_AABGCI guo_h_Page_058.tif
f9b90adabd88f0e98a3e526a594eb024
7d302556879782f39a47ccbd9fa1b644cfcd526a
19841 F20101123_AABFXC guo_h_Page_108.jpg
1a4d3954d3c4b3ca2a5cc1c910b39b2a
98f135221c7f61b8a82941d87e9092c2a7714563
F20101123_AABGBU guo_h_Page_041.tif
54c99965f0a25e4d99a4523446572766
ad1e61eacbf27a4c013d9b859db507cb1eaa97b1
50963 F20101123_AABFWP guo_h_Page_093.jpg
8e24967bbbb80e4448a689d380f7cf8b
91b10bc0905eff50f3463b2b5ee800d441d2d821
F20101123_AABGCJ guo_h_Page_059.tif
49dbbf54d656bb321be7ec5432c7628b
1ade28127c0785e9f80cabdeb189033bfbf1c90f
73063 F20101123_AABFXD guo_h_Page_109.jpg
5f752ca184becfc38519250559fa67ad
66d3ae6acaf67912ea63ea5ceaf562f66eab0607
F20101123_AABGBV guo_h_Page_043.tif
81926d8a4690d350f8affa07d4ce9587
4a8e2f85b100c0cc9f9d18b8d76349d7eafb939d
81804 F20101123_AABFWQ guo_h_Page_094.jpg
120d3ded9a26860b429fdbc846e629d3
63a0effb27565fdf419c73b7a833f1021078a420
F20101123_AABGCK guo_h_Page_060.tif
fbd38af86c4ab608a7c9094670b60493
d51cd7561dcfe8d610d01b0f1439737f7d8e8493
47831 F20101123_AABFXE guo_h_Page_110.jpg
cc2f024be4d2111481e166a3adf5e75b
dd728da1b4dd09d9e25206d1b73eae14c541cb5d
F20101123_AABGBW guo_h_Page_045.tif
b9b0207331aa9f68f1b611bc96eb1d2d
d901e9e7f9f52b4be1253b09b90ba6c34b39d43f
58392 F20101123_AABFWR guo_h_Page_095.jpg
c53d76f2207872fd709b0c0a0f0ca070
531501a09d45f50f105cc1d4917abeda41b85497
F20101123_AABGDA guo_h_Page_091.tif
da1373c9d8bed1389f6f52f61e49b193
c5039f6c316a2b5ab2bb60f8aaddb4e02d756e10
F20101123_AABGCL guo_h_Page_064.tif
111fc4aed53b65235543181572f42e5d
ab882fbb150bc881315801a53d8249b32d5e493a
95287 F20101123_AABFXF guo_h_Page_112.jpg
3aff8ccd46ee4e9d129a47bba516dec2
81b178b3ee4110a78f2a9a435d3314dd87548a1c
F20101123_AABGBX guo_h_Page_047.tif
bc26dc9d68667ac00bb184f2bed5c4c8
f9e174a1c6d8c29448212a4a19b0cda81f02c4fc
66593 F20101123_AABFWS guo_h_Page_096.jpg
75111dcfc8436a677d9b4020bd7150e9
de85462906a6244555036f5b11505ecd88d103db
F20101123_AABGDB guo_h_Page_092.tif
ed3cd149bfe12c84e0934bff9492cde5
45a68b45e169caff40af48476b2b1da3e1c1a391
F20101123_AABGCM guo_h_Page_065.tif
d3403ea3ef4bb94271512cb19ef149f7
8af76380e280708b96f397b85578f89d8798de47
91857 F20101123_AABFXG guo_h_Page_113.jpg
ca803aeff92ada232c45e694ccd9b381
3b98acbf8c91d23e203875ae7c463aff889e3b75
F20101123_AABGBY guo_h_Page_048.tif
6507697e5e026049532955f997f47658
7927d4182e04d22bfc1179bce47be3bba78a7f8d
31254 F20101123_AABFWT guo_h_Page_097.jpg
2193f32ddf86549c8ab4a790439f6bf0
8db1c540259162e272b0091fd82f5e280543f11c
F20101123_AABGDC guo_h_Page_093.tif
7267c66ee271b0ab18020f050d6fb447
4abec3084bae6b7a5b80d5cf7ee336cbe1dc613f
F20101123_AABGCN guo_h_Page_066.tif
e62b2a6da43dc1c6bd91229e0276efda
8fcc5f134dd295b3569e9afe59ce2691c3dab818
93884 F20101123_AABFXH guo_h_Page_115.jpg
7e85ef0ce704a21092d7e567ddd408bc
eb1efc5cd23665eb83a855ec25ebc19c2fd6a968
F20101123_AABGBZ guo_h_Page_049.tif
b8626a0c5e702b5c61427268c4327b9e
5e03d37810ef78f3963d3575d4c22759c7cdb27f
17985 F20101123_AABFWU guo_h_Page_098.jpg
9958713d0485ab95deb5b07243ac50b2
8d11373b4c7fd3c6893e33e21029468a71cb67e1
F20101123_AABGCO guo_h_Page_068.tif
8c204ef2b50fa13badfd3f0dcb6d7284
11d34fb9fd53c12b3f213e416b143c0ec287ac83
21835 F20101123_AABFXI guo_h_Page_116.jpg
250cf65acef97e4cfb09f10820a8955b
baeb45b44d5e4bbc45acfc3a3401138fba153d08
21630 F20101123_AABFWV guo_h_Page_099.jpg
3ab07729b04f46f6ce2a871db6a448a4
05ad93bc28d0a1eccf96b96f5a4133f6be2a4e4f
F20101123_AABGDD guo_h_Page_094.tif
f335ae250e9b5bfb0b78ffbd0fe6d0ff
895f361f5215a645cccd01b611bbb8d67f7d701c
F20101123_AABGCP guo_h_Page_069.tif
cf8c2945c47548703c107b16aaeb3ef3
1ac3f5ddf5986e304953b296a79311b014e6b145
243594 F20101123_AABFXJ guo_h_Page_001.jp2
8b7bd27f460fa81917c31468124809e1
7589ed0f3611feba038e851d0e1cb881e4ac7e78
55795 F20101123_AABFWW guo_h_Page_100.jpg
56b9a81212540456e11b63e1a135b083
b0b05585b7325275b40b742b049a2af43b961a85
F20101123_AABGDE guo_h_Page_095.tif
5a7acf20ff322a4f25a0151e47cd83bb
c4456c1fd400bdee6cd81aece72c2e92cb71a018
F20101123_AABGCQ guo_h_Page_071.tif
e18b2ee6ebf75f62f4020008f1c70408
353f0d089c0c522a06be727ea2ae4263f31872b4
28825 F20101123_AABFXK guo_h_Page_002.jp2
4ddc772654c5c254d688eb5e3a8c260e
c7bea2f3de574bf0bf09e2d778aac5aec122262b
F20101123_AABGDF guo_h_Page_097.tif
cc2074310d1b5f1b9e3da84eea01981a
f7ed6bd7fb32119db3458e3e6d518375923311b8
F20101123_AABGCR guo_h_Page_079.tif
cf7573b6d1b971e7b667908774a169e2
71a1b695e8d713a47323ad5996259e8017c159a7
956827 F20101123_AABFXL guo_h_Page_004.jp2
b23ceeb382b85586d46749f26b739c8c
31b14da4b3c7d4aaab9ba2707f98a2f659c7479d
24999 F20101123_AABFWX guo_h_Page_101.jpg
9ec21d6410f3964baf677bd3d9fd97bf
8c10b7383ffd69ba8a74ace5ef860848cf8fc5c2
F20101123_AABGDG guo_h_Page_099.tif
82938a439614dc00e3d0b721df59705a
bd33863c3a6016f5ceef9516a886b412c105e340
655705 F20101123_AABFYA guo_h_Page_024.jp2
020352d96eb96eca383368e37b6477a5
8511b182f9d2bdeb5b9ce34636f22cde3dfcf8c8
F20101123_AABGCS guo_h_Page_080.tif
d9801cdce51ca6a41945d65943202abe
f8d086a57e0e26249f00fa40781bb353d5742f88
553111 F20101123_AABFXM guo_h_Page_005.jp2
af77c290d95537e2441c804390989690
ad686292d27777db7886a1928b8c49f8330996e2
62463 F20101123_AABFWY guo_h_Page_103.jpg
1d76512f10462e08932d89c15dd560e1
f78f422ec3bd448adf12e157db052a52f78acc1e
F20101123_AABGDH guo_h_Page_100.tif
77c50d5c7bb14f866e261236846dbc18
8f7dddfc66afcde626506280e08f41701efac998
1051961 F20101123_AABFYB guo_h_Page_025.jp2
a047bd88f4c5e050cdd84680ebeb26c2
a2a47955700d1fd8992b6e7c29fce4bedccd9a3a
F20101123_AABGCT guo_h_Page_081.tif
ea0db44c3792abdc710f3df7107bca09
cd05cb67d9ef860169975ad9cbe43f174d292221
1051980 F20101123_AABFXN guo_h_Page_006.jp2
067ba1ae171373cb19c9577434c8adeb
1ed162c0bc3b7dcb232e05bbe135e337a7fe3600
26392 F20101123_AABFWZ guo_h_Page_104.jpg
281ad134c06c5e37f6e75132236be931
c02997503b303fa8765a9e102a21f9d510e6adbc
F20101123_AABGDI guo_h_Page_101.tif
4e3515d5f80b82a88b3c8630af1535cb
080a1d9b96f3187a081230c47616feb11ba0ad9b
278745 F20101123_AABFYC guo_h_Page_026.jp2
760afff1ce26dad8484cc831d35fa96d
c776abf31c3a15c507152254b6fef7f7ada34173
F20101123_AABGCU guo_h_Page_084.tif
07e8db9973ba126f2023b59c03baabe4
0f9a0278630ebda9226c4c50b165dd01eb8d33c9
565134 F20101123_AABFXO guo_h_Page_008.jp2
bd18e9ae0e616d4683497c2bde2e58b3
73a2dabb95f45cb42ea1ece04e7cfe94def5ca54
F20101123_AABGDJ guo_h_Page_104.tif
ef19dd58c3572f6754666599adde57f4
8cdb2850e7bb4956a4c9418600bfee9a9884c38c
895649 F20101123_AABFYD guo_h_Page_027.jp2
89318d3bd3f4ec0b87dc0171c59669df
604f856d24a4237c6d2cbc413b918ac8c48b32f9
F20101123_AABGCV guo_h_Page_086.tif
d710a2d793a6e44fbebf52ba8548aee8
0e5804fe040aa5bf6ee0ab0bc9e6d587444a3a61
1051981 F20101123_AABFXP guo_h_Page_009.jp2
fde1ecda0fdbcff7e8e8a8b71b5ebb3d
201ac23a84bef45591ec0de5c6ed3f2f31072fc9
F20101123_AABGDK guo_h_Page_105.tif
35750ec953dab20927f28404a4de8359
dfcddc483e9e63aa1906e512a0bb9443fd5036ea
1051959 F20101123_AABFYE guo_h_Page_028.jp2
3aafe0b155a96a733692ae834185a342
1d7c0a4b0b916aa1d28d15b1c6aaacf879eb6f4c
F20101123_AABGCW guo_h_Page_087.tif
78c4fe4dbb0baf66eeeaac2cf5df0f1b
ddfe79eb6f2934fee3d06717c8db54467bd95b03
1051948 F20101123_AABFXQ guo_h_Page_013.jp2
a904dee49ea67054cefe174bbbfdde64
63d7e9b7041524ca1073baf1f3d062c74d6c1b6d
F20101123_AABGDL guo_h_Page_106.tif
d705daac97df6102b82fdcf8b3fbd6ec
874717601e83d106e2996948d145409d3276f825
1051912 F20101123_AABFYF guo_h_Page_029.jp2
af7f7816c653b3eb303c2c0f0bbad095
3ad08cf939f6e9b4486e5e54835bdd1aa57c2fbf
F20101123_AABGCX guo_h_Page_088.tif
9fe7867aed28bc068952ecb45ad726ce
9f114220b1b6a6a0adc361f417daf355b39bff53
1051976 F20101123_AABFXR guo_h_Page_014.jp2
5420d5c985764375464c7bc047e646c0
810d6d4ea4a03c651cb83cc966447324a209dbb0
47505 F20101123_AABGEA guo_h_Page_009.pro
4fa704201087949108ff3e68ce034e76
23a5154f18b08cb372b1194bb2bd3df02b600ad4
F20101123_AABGDM guo_h_Page_107.tif
0ee60d3e265580cb2855c21f34cffb9b
c142988ecaf3c6fc47853c1f408ff4b1ca0afae9
928718 F20101123_AABFYG guo_h_Page_030.jp2
0850c46985a5666fd2ae0cb21b107f3c
7edfb1b12587c73532b58cd44e81e38c7cbdf0d1
F20101123_AABGCY guo_h_Page_089.tif
4d3e5c1bf59bad21c8ee3321a60481f7
cb4226ae9ea5a5e7398ce5441ecab2ef9455413b
827190 F20101123_AABFXS guo_h_Page_015.jp2
dca38f812a2a151a502cec83a9d08742
5885cf63cdd551998dba91547d34c0d42176a31d
42286 F20101123_AABGEB guo_h_Page_010.pro
69f0728491436792d1d78fa52de47b0f
b2df4335988669689a3755832d2d7806360e779a
F20101123_AABGDN guo_h_Page_109.tif
491249cb703f90dfcb4122e7d3a12896
ce3294b6c6f44439bccc95c1cdd506ffb130ed70
517530 F20101123_AABFYH guo_h_Page_031.jp2
672da58c268da6f7935cae41e3db856d
22f722c320fc288eedcd56997aadae9a10df1417
F20101123_AABGCZ guo_h_Page_090.tif
f304bd4b31d9973708c71dec0b98c07e
968b9b9250913f5a3a6d225a06345f54b61b00a8
691590 F20101123_AABFXT guo_h_Page_016.jp2
59ebebc5252cbb9ee7f1c99fb32b392b
f74f978f322d1521c80549608699cd26bd92baa2
55611 F20101123_AABGEC guo_h_Page_013.pro
78634600c58443d80b59bcd2b8ce27fc
96a7b34aa4fcc61f6878d45dd7f192954f3c5f6d
F20101123_AABGDO guo_h_Page_112.tif
3150719f903d77487cd1aee7773b17fe
38d342642e46c9bd9c84c0c9e5945fd740c6414e
832988 F20101123_AABFYI guo_h_Page_032.jp2
5d7b54253aa64685f5a23a7e5d2936f9
c79c085670e765d0550fea78827ca0ae1b58dfba
1014880 F20101123_AABFXU guo_h_Page_017.jp2
8cbd5838c38fec186a6fdef06bff23e9
b58477027e9bb243626c22f9d2f51cd00968ce03
37843 F20101123_AABGED guo_h_Page_015.pro
2bdd0b0ff8b4cafd491bbc1c4af07459
15066cc2801f756063cd603c91d1ec5d65132a63
F20101123_AABGDP guo_h_Page_113.tif
8d6fd310ec51b090aacebec45b612445
ebf97c53780d64563b340f4ae74345e85422c63f
556167 F20101123_AABFYJ guo_h_Page_033.jp2
e6364cc182e4d7298a11550c3b45bfea
3678861f2c4530d2da08046cc272a0fe434f2d7b
955714 F20101123_AABFXV guo_h_Page_018.jp2
157bcdd151b20e36ec823cfc8b87882e
80a8d83903a6a953496406b924eb0a2257dbb87a
F20101123_AABGDQ guo_h_Page_114.tif
48cc6ca9ec0a7adb140a9aed5ab3a6da
9bfabd54687944504891531e3adc8383bf4f12da
537132 F20101123_AABFYK guo_h_Page_034.jp2
0edd4de234d454ec982929478ba658f2
e72101c6a6528f89f7dead31a1a0390ca045e069
992233 F20101123_AABFXW guo_h_Page_019.jp2
ed50d15896d10726a4f72a3b940d3151
e42d511dba91c7c7c9e62e423d796e6450db22d3
33910 F20101123_AABGEE guo_h_Page_016.pro
c2f7cc7980377fa4a8326befdc5002e8
aeb5780afd709926bcfd843e8522cba9ba248d9c
F20101123_AABGDR guo_h_Page_115.tif
0e3668c3a875d6f43d1ab7552eebd958
87ee98dcfcae1bdf0bed548f56044aee3a96f155
1015924 F20101123_AABFYL guo_h_Page_036.jp2
d40b70a0a10ceded343e0f58eada3c6a
329783a9286ad0c5983ae3490f212f034b933fb8
936055 F20101123_AABFXX guo_h_Page_021.jp2
5c2542f752cffe5cae7d37c23ae6c2a0
3d8302d28c720d18c65c0f666c15bd0826f1a000
46997 F20101123_AABGEF guo_h_Page_017.pro
439bef10bc7eb9372b063cf643275b6f
e6749221ed6d5fbc08eb832c8d0e38ae463e1199
746290 F20101123_AABFZA guo_h_Page_056.jp2
6b8d31391b4db22817ecfce4d693a2ac
069f4cda1089104d1c44b0ec69d11d4cd40568cf
F20101123_AABGDS guo_h_Page_116.tif
9c136717bd67d14d052c2b8c3ee8b20c
499036ed4b2aaf41e3e3450aa64408633a511495
477605 F20101123_AABFYM guo_h_Page_037.jp2
7d7ca98a182628a7c6bad6b7851560a5
5d7c2796312f0f33a42e34cd01f36520440f5a6e
43711 F20101123_AABGEG guo_h_Page_018.pro
b5cc431dd03941eacd42fa2c524aee71
fc65b8e525cb61d32c86ca3fc9d8799744f2155b
983877 F20101123_AABFZB guo_h_Page_061.jp2
61c34334009b1b9c7aaea486724ef7fd
498495e4a79b70cff4bc590520e7e254037ee007
7844 F20101123_AABGDT guo_h_Page_001.pro
cc68165e4996215761c4c4802b2a5e79
7b71c871448f03a737664beb28613bbf8332414c
915909 F20101123_AABFYN guo_h_Page_038.jp2
1ba2e7f682c846380a7b857e0956da86
a0ced4f162dd86064fba407e484a75bc9b53f4a2
1051897 F20101123_AABFXY guo_h_Page_022.jp2
bce0e8e5a1285c71a9f006175a4b5757
3873c760596ab6ba78b16d044be4ad411e9ccaf9
44690 F20101123_AABGEH guo_h_Page_019.pro
44b9831775b49db4e0f41a5de20d2350
5977f851eed1d95dcfb0bf03982e781b1cc6a041
455978 F20101123_AABFZC guo_h_Page_062.jp2
d37c9e11f85f220fd2e65a38e473a337
6c829da8ca5fc6539d0a72a28f57809286dc602b
855 F20101123_AABGDU guo_h_Page_003.pro
2a7b6114bceb52e96e3f509ed00045ae
f9c00bf140c0d594d6be8c215c43aa2aae161877
F20101123_AABFYO guo_h_Page_039.jp2
3c18ea36325895233805e9228e40c3fe
e0b7895b8e7b6092ebd7f2c58094d158b39607d8
1051972 F20101123_AABFXZ guo_h_Page_023.jp2
458d5a000a38a3ad20eaff6e912c3c8f
b4e5d77e16f75e4f6b7ef6d38fac0513fbfff40f
12805 F20101123_AABGEI guo_h_Page_021.pro
54de3381979ac11d774fcaaf1f2353ca
68da2e52f7e279050442a8fcb3cedae6ea5340cd
798273 F20101123_AABFZD guo_h_Page_063.jp2
8e875cc18fef7174bdcc2826f28fdef0
e7a7edc3402d3433a419a956f7626fd9c82337aa
43340 F20101123_AABGDV guo_h_Page_004.pro
cbca1c880abdfafeabb1603e37ffe157
1bcf931be0fdfd107e958b16f7b05822052c5a8b
1051956 F20101123_AABFYP guo_h_Page_040.jp2
96eb1c124aeeef2db6c9fbdc48ea50b9
445eeb0db19a18f184dd9650d30dc5d0b9cb7ccb
52538 F20101123_AABGEJ guo_h_Page_025.pro
00e23971b47356e6983b948686aa4376
51a4cc7faa0397a6fea6948b739fefddc3ccef4a
789014 F20101123_AABFZE guo_h_Page_065.jp2
18973a777b49483cafd109a6c276411b
6425229336d649644fa60dae2edabf07176dbfbf
24321 F20101123_AABGDW guo_h_Page_005.pro
3f67e481b2ff41d9cefa72cc03debaa2
d72f0a9ca035208fd843b49bd828ec1ee954a28b
764018 F20101123_AABFYQ guo_h_Page_041.jp2
c99e9c4a3ad4482d4499246993e0c8b5
7e6c413f8d2e823c8cd1ef8b8c48fef8d8957952
10167 F20101123_AABGEK guo_h_Page_026.pro
6416ed8aaa76002eecdee280022c36cd
7d32d4bc875579d689ee9367e28b5a798d65d4b7
958527 F20101123_AABFZF guo_h_Page_066.jp2
6fd7ed7cd7f72606eb6dea501ab14070
2ee4d5440dcf029afe36c978374e892e42bb8573
55859 F20101123_AABGDX guo_h_Page_006.pro
89f5147bdd24e4b6034ef27c68c7ee2c
40d40c306e3b345fc4db7dccb170da6e59095ab6
1051985 F20101123_AABFYR guo_h_Page_043.jp2
9a3a050de4ef84bd573bb52a486b3136
965a79e53a58b32a39f0c8d34b827c7efea72b45
6185 F20101123_AABGFA guo_h_Page_050.pro
e6f8749b1b72b8963949c28626714fd7
da67dafccfa461d988993baa1897f8226c161f35
30684 F20101123_AABGEL guo_h_Page_027.pro
a01ab69bf92bc6a751dce6b8ea873d84
a1829c3e830278f0415868a95b4cf3476a110928
79028 F20101123_AABFZG guo_h_Page_067.jp2
5136533141f6bc4f94d5b8dd49297b57
f1be531d3ea606318b8ca34c99152af80b677d4c
28328 F20101123_AABGDY guo_h_Page_007.pro
c5440e8c196bea6f8ac6157d0414b053
08c60cce52ab18c2b3bb0ff44d77f5df0cd73c76
733991 F20101123_AABFYS guo_h_Page_044.jp2
63eebf6a0205993cbc9cd18bc408aa84
d61fc5b8923c0187575efd9ffa6cfa5da71fe52c
45497 F20101123_AABGFB guo_h_Page_054.pro
e3010ef6dd7545d93ebfd05917a6db5e
7c05067dbabf6779ca1bcf49922cce2464b44798
51367 F20101123_AABGEM guo_h_Page_029.pro
cb378b5a3d22fcc6d42c985c1c17a41c
1008bbfdb2a2a4e7df1b39c3cd3ed0bbcb928f96
61125 F20101123_AABFZH guo_h_Page_068.jp2
a31632043f23a8e87c71575a9d0bbf93
7a65bb6a9fb405779fa3f7ecd51ac78f8a223976
14174 F20101123_AABGDZ guo_h_Page_008.pro
4a4ad31f6189e5e3d05b5bfacc0950ab
0cbfb576463542ce29e7d4c677115d87e97d96f4
661333 F20101123_AABFYT guo_h_Page_046.jp2
700c078a0c3a7c86fbb854f6ef87405f
18fd5fd66db7e8edb10d083f9c039e30b2c9e174
25568 F20101123_AABGFC guo_h_Page_055.pro
0cf422e696c2ac2a67f13267aa3d2885
5d067216f3980ba2ee5a9d4e33deed851dcf70e7
43888 F20101123_AABGEN guo_h_Page_030.pro
c6a44d9fa0606e2383ab3541cc3112ec
2f49ce732bf792c4fc2e19e98fe4d8d2c85a10b1
738787 F20101123_AABFZI guo_h_Page_069.jp2
a04fec9356251ae4a5484b2655943fda
f966cad2970881b727c5a65bab1f346e660b9395
526555 F20101123_AABFYU guo_h_Page_048.jp2
432c37d38b2d6d1ec8369a05004be25b
382628cda601cce72f6981dbc0ee9ee7c791a754
53557 F20101123_AABGFD guo_h_Page_057.pro
2aca1e87321330df5bb0f7152dac8fd8
bf2f6a16f554885dea9a01bccb94c54c1ded1899
23573 F20101123_AABGEO guo_h_Page_031.pro
3d81a28929a1c5d5f9135808b09b6a8a
7046dadde60aa3dedca6ed7506850cd6eea43313
1051947 F20101123_AABFZJ guo_h_Page_071.jp2
9f3abd0ffaf56556e205a109539e51c8
7766f6ef64070b50488bbe453c2dd3e9a9277023
982270 F20101123_AABFYV guo_h_Page_049.jp2
f100a1707b365a9db79f7c9474324f69
a37539c57bcdbac99ee18efee9ba25b57db65337
46630 F20101123_AABGFE guo_h_Page_059.pro
9a60cb2b9fc96c6b67492853b35bb68c
0ee09040cd41e2571e9140efc569f4423b4960b4
37914 F20101123_AABGEP guo_h_Page_032.pro
e74a51842b222664a1c4a32d81d086eb
a6a8d5a0c4ff138a4278ec9d9662f49fef86aca0
F20101123_AABFZK guo_h_Page_072.jp2
467555f1c3fa18c90b813c88fe328e73
c1255675ab5412df35c4f31951273cff4f567448
1051939 F20101123_AABFYW guo_h_Page_050.jp2
89a3bf8780ffe44ba006aeea6a8f82fc
c0677471a3c59cc6ec30c14daf02b60f832b806c
25080 F20101123_AABGEQ guo_h_Page_034.pro
38b781c87ee719c30d4ebbfa35f6a5a1
bb1cc76428dc43a9ac955e5d89d15c0dd7dec321
905954 F20101123_AABFYX guo_h_Page_051.jp2
6f5824bda5f0d2cf0b352dfbb3a63d54
031cb740119228afa46569e3f9f321b66b23f93b
45847 F20101123_AABGFF guo_h_Page_060.pro
9d19d3572b63d94e6c7ceea97174dd5d
373bd67a9696b2f63fd150a5949b94788b63ae88
46773 F20101123_AABGER guo_h_Page_036.pro
5dfb8e4b3abef52020b9f40d3099b319
4c9f285cb468e7206541cab416804d67cdd4ed56
F20101123_AABFZL guo_h_Page_073.jp2
6d44ede928e7d8b34a9965ba4153ca6e
e139a9715cf8bec87c9bd4406089f5079b8f1d8c
F20101123_AABFYY guo_h_Page_052.jp2
56fb91cb8255c292a5a2f61d32f97861
87825ed53574e3e0d07faf6312925690ea51fcd0
46328 F20101123_AABGFG guo_h_Page_061.pro
123182776a568ac1414809d585bf2b1e
5aa53434996d2b9b24b1eec715dbcaa3817cf951
21934 F20101123_AABGES guo_h_Page_037.pro
766eafaaad575f0939046f80bf947c48
4aed33e18f5e6e128b555b4117bcda463dd73109
881641 F20101123_AABFZM guo_h_Page_074.jp2
a3e64d882c89594137e7d958c31506cd
b5e845d2a7370df7529b8d20158a6582e99889c0
19158 F20101123_AABGFH guo_h_Page_062.pro
a2750303b7e73dcfb9ea700a3a545df6
292f5e70dbdcb648eb7a8192260d757ba8c46b7e
41645 F20101123_AABGET guo_h_Page_038.pro
4e1fb8cf767bf5979e232c0b06c5b3b2
9dd52dee0e21048a227414df0e8e3889f9fa40b9
329447 F20101123_AABFZN guo_h_Page_075.jp2
2f4db3fcf1962d5c9b362945f832bbb6
5c5d9afba44a5ffe3e058215b68b3c96392581ad
1008164 F20101123_AABFYZ guo_h_Page_053.jp2
a2aea1b601cb2fce4d252f7c58872f93
9f5204c10fb9d76fe5b0867427ff81daee76f8dc
35847 F20101123_AABGFI guo_h_Page_064.pro
58dfca02309ce288f07b55fb04211136
7740ed268c4625a8f062b878d5fe860104ae6605
46413 F20101123_AABGEU guo_h_Page_039.pro
e49209e81bf4de6496c60323283200c8
d321599e1105997c2da0eb35efcfbd5d40d6c12a
20579 F20101123_AABFZO guo_h_Page_076.jp2
eee558a447dedbb8aaadff6f498d06c6
70bafca02f42d813c573d311ca5dbf8bb9424d7e
35292 F20101123_AABGFJ guo_h_Page_065.pro
c8f2f80d279dc6950e7a60bca1d0784a
8167bbf48b61c84f0042c3fef70a64132503e6a4
35902 F20101123_AABGEV guo_h_Page_041.pro
977981ea0c7476939cf2e2f04894d8e5
6bbdb20dfbfed86325b6a3e1c382b82fbe9b17d0
24615 F20101123_AABFZP guo_h_Page_079.jp2
c459dd8c5c83d3d5b71d94f459c2abfd
292a8197bdbebad3b96aa29b1a14690b0120a0a4
43204 F20101123_AABGFK guo_h_Page_066.pro
329a2144a746804f0b755bff8a22924e
a2970ec40dab3ed9f6e076b205dcd78352bd833d
41284 F20101123_AABGEW guo_h_Page_042.pro
35b0ec4cd9e05281628ad33cdbb11b3a
42972b9166b0febef36b5b144905741ed04c403d
459331 F20101123_AABFZQ guo_h_Page_080.jp2
96a62c63bce618d2499a532dddac642b
1a4a6599265cf1d65e04b055efd219a9b1b90002
40975 F20101123_AABGGA guo_h_Page_096.pro
e7dd4ef22ba18c644c819f5817d8c218
efd0e7d88342f795967e7eff24b4f31a64417b69
23296 F20101123_AABGFL guo_h_Page_068.pro
7cdc623ff3692b35c60e211db8b59564
ed8d30231149b37018aff7a705607715738ae021
35757 F20101123_AABGEX guo_h_Page_044.pro
977ed943c15d4a721ee72719c0d14f9b
784cb94c374f53c683c614c094580603102d2f60
805366 F20101123_AABFZR guo_h_Page_081.jp2
c354d6f022e624a1ed912293b299eadb
102acd944b4e32af2780b779c07f4d54710719b1
6600 F20101123_AABGGB guo_h_Page_097.pro
b39430b3dbc553f83a1628b082df558a
42a4472efd93b9c399630f113d967d8313cbd2a3
40647 F20101123_AABGFM guo_h_Page_070.pro
3187852bd5351209eb8cbb2b319be202
f34684fca1663d3cc0807021af75978952fb08af
50344 F20101123_AABGEY guo_h_Page_045.pro
20198c5984a2e08edb8c705625681f29
629cf4a9db01f24888e3c77bcf368d784f59fb09
1051966 F20101123_AABFZS guo_h_Page_082.jp2
6d4edde2b7982a4efb950620934aa1cc
cdcf114134bca2c704ee0f5f6f5b85c8cbf45fbe
3555 F20101123_AABGGC guo_h_Page_099.pro
b003b9031133f01382c4227a9adcd33b
f124010ca04afa8a3e4246fb312fbb735055bd5f
62382 F20101123_AABGFN guo_h_Page_071.pro
c1577aa35c9a3c4e5c7cd3ce8807505f
47e263f2b91c0817134532e5472b31bae65067f7
40120 F20101123_AABGEZ guo_h_Page_047.pro
5091ee48667b8efee988484411b8274d
5165eaeeaf2929506a17a408975640b7c1342507
1051957 F20101123_AABFZT guo_h_Page_083.jp2
ad0a4f57fbe9bff56e4fb34af7bb9bdd
19a312c44934f86ee99378c2493c2b4c8e62cf0d
2415 F20101123_AABGGD guo_h_Page_101.pro
8806fae744d09f5aae490a30c7e5a3a1
b64081c507fedae16def0e628179ca1985ce496f
42028 F20101123_AABGFO guo_h_Page_074.pro
ef3ca64eff9cabd1bd2b883e9520323d
8e8b9288209c30d18af143ff536543aeee88b2f1
835070 F20101123_AABFZU guo_h_Page_084.jp2
35a7881cd0740ef288b12bc35c965e5a
7dccd8315284d9546856b1a5ae0f1da28a2f35b2
5841 F20101123_AABGGE guo_h_Page_102.pro
4ac994d34c6c4b1e0ba0f7433bf74ce7
295c372e556c2f4c4da4c85b4171917fffaf2794
14213 F20101123_AABGFP guo_h_Page_075.pro
89d802440e53e758aa4b51da13dedc2d
89f96c700f0a3c52b6c2ce745c1efcf5b5e706a0
768594 F20101123_AABFZV guo_h_Page_086.jp2
441f2b60c21305825ed7f0c38f529905
79a5de565c2acf023d19aa14a4302081c087c3c1
2285 F20101123_AABGGF guo_h_Page_105.pro
c680c7068ed40d438a2a4a315fe95b29
e2f0e394e7c074f68877a1f6280cd26f3fab9d96
6689 F20101123_AABGFQ guo_h_Page_076.pro
a0535193b915428cdd18ffd21adf1287
83326767a5f9c2f4ff25feee9bc96cbf44d6e8b7
1032400 F20101123_AABFZW guo_h_Page_087.jp2
8d5ab8395c0860ac30e4dd987e881ff4
76d3b4f2fa6709d5b1b23785f7e0c49b44b44753
1921 F20101123_AABGFR guo_h_Page_077.pro
88bc60e186593f01077add60fc31dd13
9508e1fd77cc1f2dda987bf816fb6a5facc26e10
F20101123_AABFZX guo_h_Page_088.jp2
aa6ed582b408de54ed739b9adb83d39b
6d74dbdb00e379f08cddb3f59f23054378aca7aa
3171 F20101123_AABGGG guo_h_Page_107.pro
f57f53deaab591667b6e964fbafa5e5e
d6aa42ef603dc482953fb2b3bbc4fb571e8dd62c
8875 F20101123_AABGFS guo_h_Page_079.pro
c8cfc2f452c745d5d2f1ccdd016cc82c
ce046b6cb8d7d493770c35bf36eb75bdccd5e3e0
998453 F20101123_AABFZY guo_h_Page_089.jp2
7a02772ed406677e85a015f2c170f263
ce552658900a53e2feefe9d336c91ea5cf175268
4105 F20101123_AABGGH guo_h_Page_108.pro
cb6805f0b5afb3d7a84d4e560f6b44af
de8e75f9ab2cca71340bff8cac69b68acb2a1508
45470 F20101123_AABGFT guo_h_Page_081.pro
164b9557fb4d5859dab00c88e9ca2449
d3ea6c445e456980f4368291a81da4a58338f58e
924646 F20101123_AABFZZ guo_h_Page_091.jp2
e01121f19846487eaba5fc82ccdedb3b
cf71e4840f6c3929eec1539697e3f0fdc0745313
36904 F20101123_AABGFU guo_h_Page_084.pro
b5727b06b240fd89493ca80e64c5c9e9
204fee8ecb731bf73a7a8c41a09f7fe7d9065ba6
47110 F20101123_AABGGI guo_h_Page_109.pro
70c53ccef8c67cfb864a1923ea01c5b7
6b57090a814b7ba641e1a84bfc948f98e90632fc
36522 F20101123_AABGFV guo_h_Page_086.pro
10e5b6d1792c4f34528904560a21bbbe
217569154f20dda919e9f2e3aa919a49fee3009f
29120 F20101123_AABGGJ guo_h_Page_110.pro
6ebdda0b0f1fa7a54b5389ebf1d12a5a
877276b67b0eff103fc0a3ba950e3e0e13f3d3ff
46922 F20101123_AABGFW guo_h_Page_087.pro
396643337c9c5d69d1a30fee17e45b2c
327ff100eee336cf084549d05c00df8b205fa048
49493 F20101123_AABGGK guo_h_Page_111.pro
31fdfab035b1323a1b5e756a13dd057c
2ab979a9c32295cc17b1a195bf15bbf12c721e9b
31870 F20101123_AABGFX guo_h_Page_088.pro
6783a0ea42421dbcbba8b8b4f8b938f4
76bd1ceb49215c92b2b1831ed7b865673bee7bb4
1868 F20101123_AABGHA guo_h_Page_017.txt
2d99089abbf7d25eafa9367234b36bff
ab2ea2bead5fce6de4149a1b9fbb5f68cb47167c
59115 F20101123_AABGGL guo_h_Page_113.pro
65a9ca2c79f09a3d21b9263f9dd11dbf
800ab3ab79336a100c622231c13781fa851c76af
45557 F20101123_AABGFY guo_h_Page_089.pro
e8dace710f73a68b71205c67b35de655
b2f39582020d8aa8beb4163dac3612f5fc2483a0
1748 F20101123_AABGHB guo_h_Page_018.txt
8897c7f1680845213fbeb223d3b6c076
242c429afdfc6da52a368dce3f91188aa05a79d3
57352 F20101123_AABGGM guo_h_Page_114.pro
fa278154e353509fc59e308980b199da
12d559a86f35fcfb0da270f7c43a463652f35801
38259 F20101123_AABGFZ guo_h_Page_095.pro
3e96d76d685d09662a30a66c3349516d
63039df4549dd8e52fce5949997d28d163e008f3
2182 F20101123_AABGHC guo_h_Page_023.txt
70419f84502b6b58daf0efb2bd21ab19
a29c35490da1e420d27290f5282be50e81528d46
60947 F20101123_AABGGN guo_h_Page_115.pro
1573c8ea72408e3f7b683c4713beca7a
19b723814826e5016b16fbcf3f8b25e6f8980df5
421 F20101123_AABGHD guo_h_Page_024.txt
56a1ad43ddd426249a2d4e1626705033
5c2ff14d2a63c2f2f3f24d94d3f464aae7c4964f
10005 F20101123_AABGGO guo_h_Page_117.pro
8bb92c23a16a0536c547e7370d75dc18
d4499f8948bf19f8031be9c1300abf43733ae3cf
532 F20101123_AABGHE guo_h_Page_026.txt
752f67128f6724140171be5c33a90b7e
de511d877ab8b92ade317f32b39a58f90a32b638
422 F20101123_AABGGP guo_h_Page_001.txt
f4f508286214ec2e135033923abd6672
f22f8a68c2ea8fe24ecd15908072c8f1fe34133c
1512 F20101123_AABGHF guo_h_Page_027.txt
14a90158414f982bbc984f61340b11f5
5b9b4e72035e2a8a237b952787c9aaa72e250374
105 F20101123_AABGGQ guo_h_Page_002.txt
5f20a23e20d9dc57b5a6f60f5ffa04c2
e3c2b043a59d02cd69de85c8c65e2fd0581878a5
2023 F20101123_AABGHG guo_h_Page_028.txt
123fa365c8bef7da17508ee171c5618b
746016029d91e231951b766e7ee21a7fad2344d1
51 F20101123_AABGGR guo_h_Page_003.txt
1bd1e88db54346dd09be50e1dcb17cab
5289ed4b659f8ea615947e1f5a3b843281ce18eb
2419 F20101123_AABGGS guo_h_Page_006.txt
7cbdec455f5adc34eb740bc704b13660
191c1447b1e86c991ddff319660b6eb2f0587312
2086 F20101123_AABGHH guo_h_Page_029.txt
978e8d603c6676d100b879318df333f9
7db1b7c04f0a446e12603f3742f8d080b0726aa4
1226 F20101123_AABGGT guo_h_Page_007.txt
6bced5f303b106d93ce3b8e9c10cb728
ed81439f68c358ff0ad4575d64585b2a8b6be3e8
1905 F20101123_AABGHI guo_h_Page_030.txt
b608f5f052a4846af8834c2388e88d42
53d7dd424016e0219e7e4ff225db3b693e93137c
660 F20101123_AABGGU guo_h_Page_008.txt
8fa15d49ff323d83dec3b057bdc21505
4d5aa93384fdeffcd1373184eac67d68a20ecb56
1180 F20101123_AABGHJ guo_h_Page_031.txt
24519389c5258b17f2c32a27f14f99bd
67d3587330c2cdf70cdac9d07aedf5b09f101f3a
1707 F20101123_AABGGV guo_h_Page_010.txt
050f9a6ee92dee3f96aa977e09690f14
335592186a181f9977c986b527753c1b2cafdc75
1745 F20101123_AABGHK guo_h_Page_032.txt
6581d618653297ae9bf41fe1b68c2e51
95710f63e3f35ddc6c6e75ff51df04c8c3d5b6c2
2036 F20101123_AABGGW guo_h_Page_012.txt
5ed4ecdd1197ce7288a0db19df03e211
88b480df2b3e54b1fcb8c6946d7845c77185d3de
1829 F20101123_AABGIA guo_h_Page_053.txt
e68e4242768e14628dacc97b1c0376ab
c66e46e41338aebd91d07a0b7b94099667214ca5
1292 F20101123_AABGHL guo_h_Page_033.txt
bc6c294e4110e3a8b7406247d5632228
6d89dc794c88a10484227b4b587e4ac77e9d857a
2191 F20101123_AABGGX guo_h_Page_013.txt
b8a132f2c5caf2ea4813008cc1125d6c
3209824fb357df6cf19accde7344de22c1014df6
1914 F20101123_AABGIB guo_h_Page_054.txt
72fbadf8b815d9096c6b4d25ed4ea06e
868e62411c1dc5a114f158fae47e65790e503c1f
2010 F20101123_AABGHM guo_h_Page_036.txt
adb534bb5e0fccee6c212aa3f8434c48
1eafb872e785ede938842e24a2a036dac50755d5
1840 F20101123_AABGGY guo_h_Page_015.txt
e5e6885bfcb8dc8614d00ce00104ae9a
281eaa9a48d4471960c32347ebf669c943c69ac6
1068 F20101123_AABGIC guo_h_Page_055.txt
328c69b4b26c05ab2fdc03d5c830286b
024583501b8c33c23c1b9916a58cfad0e2a29d73
1001 F20101123_AABGHN guo_h_Page_037.txt
4e7214fc2e19ef5c5ce741a3dc692265
f892fedb44207103c95d7aa61e424ebf5cb7578e
1937 F20101123_AABGGZ guo_h_Page_016.txt
b7ca28e9f07eef44aadd7fca316c76a9
90f805bd44659bf0c2ccbddea40f018c80f1a2c5
2184 F20101123_AABGID guo_h_Page_057.txt
e5b70f033563d814105c37da23fe3f98
42ca3a82859add9e418f1fc6f64b87d23be8a78b
1778 F20101123_AABGHO guo_h_Page_038.txt
8d41360482cf8839eba60e9ec64eac8d
4a6523aeee60287aa0568a520537edc86cbbcedb
1487 F20101123_AABGIE guo_h_Page_058.txt
7f7ee440d577ab7f92c7908bd81c681c
660167392132876823e17eac2905861866a38a53
1944 F20101123_AABGHP guo_h_Page_039.txt
bc0e57f4e04695496adf2ef1d7792cd4
0a3d2dc49f9e8f21745e2b8ac155a1442e6819d0
2171 F20101123_AABGIF guo_h_Page_060.txt
67bee9e0b0af8ce1ba5b2072aa66889f
2d3d6b37304c62446f70be27418e9d387f947740
2179 F20101123_AABGHQ guo_h_Page_040.txt
498642e4cf038aef66395c9cd3e8422e
6eacc18792f93fb65c82d0d1a9b1fd5f8d191d24
1938 F20101123_AABGIG guo_h_Page_061.txt
538453665aa73b15141d5be826a51340
b26395df1baefa1688dea9bc4d4c3a818bd37f29
1899 F20101123_AABGHR guo_h_Page_041.txt
5cdf36c33845ecba3389381d8f4df87f
65b3d63ec2ab9557a800e8e04e9da2aa438276b2
848 F20101123_AABGIH guo_h_Page_062.txt
4e6c70d2b56381a977ba51838a5f4fcd
435e37a936ddf9ec6ea0f47c6b6c08b5bc423f24
1980 F20101123_AABGHS guo_h_Page_043.txt
7b6cc86dbef1b68f0482d270c5a7ad12
100adff306e347c737b81a03cd2afa9226781634
1927 F20101123_AABGHT guo_h_Page_044.txt
859f9ca2fe450a5eaaba44b9a9d8aa44
48ce2f6ead083356e72e199b8da0ff6d07808697
F20101123_AABGII guo_h_Page_063.txt
ab07d5c47738b9be77108d14ed8b24bc
daa3116227d66ab28ab4b74c41d7d0060f423d4c
2076 F20101123_AABGHU guo_h_Page_045.txt
bcbf71277763d5dfe54ce8319783e1cf
eefc5b9a0d738fb91db5dc1e0150e6f9008313bf
1606 F20101123_AABGIJ guo_h_Page_064.txt
7e568d573bfe61a3d8e437e7ec4a9874
f7321806060a428db17d12eeb073bb72c873d3d7
1376 F20101123_AABGHV guo_h_Page_048.txt
367ae88d8cacb16c6eb4bcdaf39591dc
5fc4ff7efe42d271e3a94f6820a22fd8b8167aa4
1663 F20101123_AABGIK guo_h_Page_065.txt
04ba2fec480b6a641d1445ebf9089ac2
df15181fda053c161c16c1285461a56c91315f8b
543 F20101123_AABGHW guo_h_Page_049.txt
f228768a54c53f1e85bcbd5bd1943550
4686e33a6ada3447ea70399e809b76264be32fb3
1437 F20101123_AABGIL guo_h_Page_067.txt
e949041e63e212f43f186e96114acc29
5c174b8d74e430c3bb159b69896e1be828c9ad1a
351 F20101123_AABGHX guo_h_Page_050.txt
1dfdfbd97c1ff58d47cb1cd5980ce512
ec83ea7d8fb0c1c381d43d4d1099debee1234d1e
2022 F20101123_AABGJA guo_h_Page_095.txt
7891e508383bdf30add5fff2b6565bd1
bc36bbaabe3adea08c0ac5d89f89a672109bd93d
1472 F20101123_AABGIM guo_h_Page_068.txt
c844dd825b89db0fee08f1b613365815
ae44650e9ad00e71f7b39fcc3db8c551e84b588f
460 F20101123_AABGHY guo_h_Page_051.txt
f9f18fde5ec3301e558fc2283ed96164
cf8d1381e7b0c09f54b4ea76a741bf5647cb907a
1782 F20101123_AABGJB guo_h_Page_096.txt
e891e4618758cc38239ea4a6f8bd4376
164e6bb93f55b29a8765ccfeb1bf074251e5a2c7
2738 F20101123_AABGIN guo_h_Page_071.txt
d0d2217b2c3fd330c4705efc173a1b53
43cff21fc08d6e34cbabd4c39edf2a885b7ff783
430 F20101123_AABGHZ guo_h_Page_052.txt
b382a8cab221a1af8cc68e60447e1e41
69c1d3c88c9b79656fd9d0d633e7c932e075c734
356 F20101123_AABGJC guo_h_Page_097.txt
a9ea46eac48912b84a058dab52007936
1a54a14f6fabc71c60d4f241efcc61dc3a91e365
2108 F20101123_AABGIO guo_h_Page_074.txt
b99a745bbd38e48373d485adacfcef41
6df373a07624e1bb5467a778b3ba9a6a1edf4195
248 F20101123_AABGJD guo_h_Page_099.txt
7eb2ee3247786377cdde4491e42c4c5b
5637841fd89356d15973aab681499eb8abdc357e
644 F20101123_AABGIP guo_h_Page_075.txt
7f1a436be84e625d53b8f51ee20efd0c
f14f49102fa0e7411f309f3f4f6c06ab382c1857
1388 F20101123_AABGJE guo_h_Page_100.txt
060988ac95472885470415f203e5a117
99a76b9f2cfb0e65f652019c6be173a46b80ecd8
894 F20101123_AABGIQ guo_h_Page_079.txt
28d5eeb9137433fe7f10e9718dff2fe3
729d9c15a430f6830c3f9408a8994c3753aa8d5e
175 F20101123_AABGJF guo_h_Page_101.txt
d7438853dfa9745fe348660abc3f8e3e
ad8f2b96548743ea3a8ac53884cb53f249b59059
818 F20101123_AABGIR guo_h_Page_080.txt
dfd1f39fc28517952b412469c9ac2097
426261775938a92725007c98fe22066ca59ef804
449 F20101123_AABGJG guo_h_Page_102.txt
c589ef56eeaafe7b1290549b115b990d
68158854dc94907a91833b9ce93bf8938259145c
2286 F20101123_AABGIS guo_h_Page_081.txt
53f2eb1e17f559c3aff03fd933af150d
4a97166625d7a6e9f200720c25f5f2521afae5ff
1613 F20101123_AABGJH guo_h_Page_103.txt
f2a1b2bd65cde97ed4ac34930dbeb095
8346154fb39828b79699de7563b9fa60290c06e5
1546 F20101123_AABGIT guo_h_Page_084.txt
76f17bbd46fee8818f368605c3a9e8c2
75ded1e1ebc8fcccbb48680e85a10dd25b0638b2
201 F20101123_AABGJI guo_h_Page_104.txt
b6929f616fd59db684c4ec6b01b3ff5a
93f543cc338a7000f44fb09bb68f71af414fc6a6
1554 F20101123_AABGIU guo_h_Page_086.txt
5cb969994e0c3bc1e4fa76c3786b3132
8633aa43d97bff3729b8cbad181e4f484f9634e5
1897 F20101123_AABGIV guo_h_Page_087.txt
a0b97508efef25b566d7ad1452387d6e
9502bc5643a86b8d0167c5d4114c77bbcbdf1848
170 F20101123_AABGJJ guo_h_Page_105.txt
1c1d874a63e9d3eabef63be04d813fd6
aad048a66756f7cc12ee12d123f9f09558eb209c
1884 F20101123_AABGIW guo_h_Page_089.txt
60aecc6cdbfbb6ca2c6236dcc50dcf95
bb7b81da366ec6d355ea7dd73ff628eb633ccab7
1250 F20101123_AABGJK guo_h_Page_106.txt
260a8eae50e16b52bc6bf92ed4376063
073fa878ca368b63c645e3077fe5daf0bc5f7450
389 F20101123_AABGIX guo_h_Page_092.txt
45e251d46d658698297954df321f4b64
6eba6e138b874514d3112900e668332c8f516550
5071 F20101123_AABGKA guo_h_Page_006thm.jpg
63e11ee45cdc77ea13ea503f0accb994
9d6347bcf3e127889208c949a48cd16258b9bb59
264 F20101123_AABGJL guo_h_Page_108.txt
f0f373fa228813282fa362493b6c590c
8ec2af09c48dfac6d2f0c4bb02988673d3f0ce99
1441 F20101123_AABGIY guo_h_Page_093.txt
18ef25d42aa48e62a3bf617a7d662b85
8cea83877c16ae15e79d2d468ee24439c973fb1d
12194 F20101123_AABGKB guo_h_Page_007.QC.jpg
b57e5109ae59730daf087088e0e99f7d
8cf695a33640a6b78d2eb105a5ac358acbb80cc3
1958 F20101123_AABGJM guo_h_Page_109.txt
41d768b68b64362d25b16cc9fb625a9e
348cb60eca37d5dfc98b3648a135cf9eefe8cce0
2135 F20101123_AABGIZ guo_h_Page_094.txt
86ec6cb1d7250addf0a39a5a4ff275f9
7b29a83f52feee939e56a23a7875c88965b2cf50
8435 F20101123_AABGKC guo_h_Page_008.QC.jpg
33bb32f763e78fde2fd3ce26bd05e330
56ae2aa7e5ea3b9aa22860d045f2180d2b54c589
2474 F20101123_AABGJN guo_h_Page_112.txt
e99958d2ff96aaf5b094819201666e85
da9a85063647b547a6356217b40cb23da3e608ef
2621 F20101123_AABGKD guo_h_Page_008thm.jpg
e54af011a5a99c87b9cae6ce4a05aecf
29a375d3393ef6000927a3aeed18f8fe641b299b
2379 F20101123_AABGJO guo_h_Page_113.txt
358f3c7b3f2effe133d1acd3fcc31621
bff387cf09fad196fa4c3f7fed0580095cbd92a2
19129 F20101123_AABGKE guo_h_Page_009.QC.jpg
c4c21a2af596f7dbc8c8866b42bf2d7c
bd8998e9eea9f95427b50209fa1e5b74b6623d77
2309 F20101123_AABGJP guo_h_Page_114.txt
33edc3f2cfdd1e2614bc31c9b3b8f436
6866355d4cd24472042816e430cb0bbfa4110673
5119 F20101123_AABGKF guo_h_Page_009thm.jpg
64ce1f482c42152b06c71eebca75dd79
baaca9ca8fb0d95370acb4af41641be7a993778c
2443 F20101123_AABGJQ guo_h_Page_115.txt
5ffcb2900eb50578a678fcb315e36488
009b434ac4f6d1bbeae835425a2b54f5742fa211
17912 F20101123_AABGKG guo_h_Page_010.QC.jpg
5c6bd9328ee19bd43af110d72ee19360
4b5a91f68b68df5a9d6b1075db3944bd94f04ee9
2421 F20101123_AABGJR guo_h_Page_001thm.jpg
07064316fa91b7c9c87b1e7affdb987e
822794d9be9c80b5f0cc3ea887165c07ac5e5b3f
4744 F20101123_AABGKH guo_h_Page_010thm.jpg
d5293a6da29282aa13f17dd33a276517
2fc8498b53dce0230fc8ea6bdc97803da40323a1
6832 F20101123_AABGJS guo_h_Page_001.QC.jpg
465ef9fc5f01af5bf12dcf86794222ab
5bf8a6c02efdd78bab7256f82b359d5d71d0e384
21167 F20101123_AABGKI guo_h_Page_011.QC.jpg
df348358e314b5ecf7823b7e8a1f5d96
9a89528de16755366db9481c82de77dbb4eb00cf
3127 F20101123_AABGJT guo_h_Page_002.QC.jpg
fa66fc2e54e36c464c92e71950c854f7
ac6f71866e2f66ce33a1980e7a8de2bd0ae91986
23524 F20101123_AABGKJ guo_h_Page_012.QC.jpg
9ec4f0624d15f6f6bb3aa42e81c67e0e
06ebf8db98be1853587fa7de6a2f0ffc22f67e55
3035 F20101123_AABGJU guo_h_Page_003.QC.jpg
698ee1266a4e5129dfd1287a28d91348
667e0826fcdaf53c2aaf60682d7de9243002aa69
1301 F20101123_AABGJV guo_h_Page_003thm.jpg
d85f8fa13c96799f2799e4c9afd93b34
abe76c93e127127fe23329701c4d5834c5acc4c7
26206 F20101123_AABGKK guo_h_Page_013.QC.jpg
0b6a92b84a3930ab95acc40e28bebbb6
a18c1d0f08850059371c2a28e655d1eafc55403f
21137 F20101123_AABGJW guo_h_Page_004.QC.jpg
788d1929bb425061cd1c24949edfe45d
9748b65b948e434f60cf3b522c517b3440045a5e
26067 F20101123_AABGLA guo_h_Page_022.QC.jpg
22fefa85bdf0dfd005e876da1ff37f81
a914a8dbb2810a7cba6b73264284b28811ddcb61
6533 F20101123_AABGKL guo_h_Page_013thm.jpg
2b59eaaae330db30455aa30710ed2a47
fd8da907cdb1c7d5d029d4f58834c022900a0b59
5611 F20101123_AABGJX guo_h_Page_004thm.jpg
8e06bf399f16ef01ad45ef84d763b9d9
36c3a6778aeae191b1257a7500729cf0b69e7d93
26474 F20101123_AABGLB guo_h_Page_023.QC.jpg
f709fd8558d5c2332faa3967353695a2
25ef20040f983d1dded20e1d6f47c126c2116c47
26655 F20101123_AABGKM guo_h_Page_014.QC.jpg
044b2a6e57677c21b112e5b22c6dd46e
a6dd0e8d7e1931e77addccb3698c8d1a318d6394
13915 F20101123_AABGJY guo_h_Page_005.QC.jpg
5b555fb7f6a3ae2268922c3701b7c467
9af3a471350be67cda618a2c5aae20679506cb53
6676 F20101123_AABGLC guo_h_Page_023thm.jpg
9adfe123f77c833cb5f7455347f1a260
bc1276749b0cae44edcc6dd748972efa6193ad16
6711 F20101123_AABGKN guo_h_Page_014thm.jpg
3b6eedad5841b9be97e8795941161e73
c13b9d9e046f5204d428da303446c9ecfe7485ff
20206 F20101123_AABGJZ guo_h_Page_006.QC.jpg
65960f4a4e08f620b39ae50b34143ce0
b2c0962f2d7a356cb7c3fd77569c83826758fb54
11137 F20101123_AABGLD guo_h_Page_024.QC.jpg
99c6cd4939cefce3f2bab2f7d784fdcd
73ae3379bf20334022da25806ca271237df0d936
19345 F20101123_AABGKO guo_h_Page_015.QC.jpg
1428fc1808d84b931196c7a0e467a63e
a41389c222b007f171a1bb04d94e8296bfab6eba
25549 F20101123_AABGLE guo_h_Page_025.QC.jpg
bf374fe851d0211ab05d5e9d8cc4c8aa
b5c0ea94bcf18ac495d4db5c7ddcd6862e386903
5253 F20101123_AABGKP guo_h_Page_015thm.jpg
ce24903a667e5694257eed1d384bea49
c47dd62903d0b4ea3504f33fc11ba5f2eae46975
6540 F20101123_AABGLF guo_h_Page_025thm.jpg
a28f58c53ef91c0d1879cc28948a8db1
9d67b9319e60b9c92c2a0cb5b7e9226280b4348e
16860 F20101123_AABGKQ guo_h_Page_016.QC.jpg
340861649a7cddff24e894b7646aba16
da9c58d42508f75eb513367d6ba1bac10c3e79c5
4779 F20101123_AABGKR guo_h_Page_016thm.jpg
d1775d243debc5ec045a9ce4be226b13
1cc5cc3dda9cfd967a09c50b35f28777f9db7648
7699 F20101123_AABGLG guo_h_Page_026.QC.jpg
237424a982a68200eee1a9aa293c885b
fd202f7edc66cb99247482d7f7e3c95b408a9140
22577 F20101123_AABGKS guo_h_Page_017.QC.jpg
f6badf5753ea61d6587401951d601e08
40b0206e46d5310095870d7e60c8f17bd7af5b55
2813 F20101123_AABGLH guo_h_Page_026thm.jpg
f8280228a47ff4bb55b39749465f1a56
edf180a9e6d8dcdedeae187a9f5896481d332c67
21253 F20101123_AABGKT guo_h_Page_018.QC.jpg
c9cae9951781863b6515ecbcc9b56331
00acba5ba4dc0a3cd50be15b442e56db5d4957c3
4373 F20101123_AABGLI guo_h_Page_027thm.jpg
a335d7eb212d998fca57b56ada3fa0a4
2cc4b985b4af3c58665698412a5f2f286a15a696
5582 F20101123_AABGKU guo_h_Page_018thm.jpg
f8ad2b9bfa2ae492f22871d5013e3728
f8d385235261afa9981ebf8b1a36dfebf18abbc2
23415 F20101123_AABGLJ guo_h_Page_028.QC.jpg
8e913b84a78cdf8e3706bbc2cbd541e1
f9ca31fb1924f08862e93d0ec544919bdc76d35f
21098 F20101123_AABGKV guo_h_Page_019.QC.jpg
6773760214e5edab7252b15a4e794e0a
32d61726e044a2f91e7d8c37071c11bdba78558e
24509 F20101123_AABGLK guo_h_Page_029.QC.jpg
78b97c981c857a6ed32f41d00bd6262c
fb024cf0e9286bc79e798be04391d9ccdf4185bb
8680 F20101123_AABGKW guo_h_Page_020.QC.jpg
80702a8e35dba3cc4699e7d3d2eaf651
1d255f664ef073d0b07c4c4ce212b45615b5e015
2741 F20101123_AABGKX guo_h_Page_020thm.jpg
93c58ef15753fbc51c49f7111b465fc5
934243b42ae254a72fbddf5c484a05fc13873929
5060 F20101123_AABGMA guo_h_Page_041thm.jpg
e1ba95da8e22f288daadf59f0288def3
98e57fb04410c2493d470600e0225fa7c0bee58b
21775 F20101123_AABGLL guo_h_Page_030.QC.jpg
c8a1db885b3dcc98883219b84c622e32
851830e80e91fdaa703ee2a923da66e8ba1ec207
14024 F20101123_AABGKY guo_h_Page_021.QC.jpg
e1c6e6dbadbd2cf50e072f3947047aa6
7d89aceb8ccb87964e61b9e636c740cbb8a13209
5665 F20101123_AABGMB guo_h_Page_042thm.jpg
2e14ec98fb745ae6ce7b6fd2d3a4f865
4a45f32e8cd1a5e2be58589c4a7a29b05492e2b3
5706 F20101123_AABGLM guo_h_Page_030thm.jpg
a3e8a90bc664860e2edb55ae20e777a4
67ee99b9c8f97cf7db74dbb4827631388811f679
3842 F20101123_AABGKZ guo_h_Page_021thm.jpg
6ceb18953111f6206637cfafc1b055a3
0e896997b7684a4ebc68700e533b808fcd494526
24340 F20101123_AABGMC guo_h_Page_043.QC.jpg
9cf096958a64e2a8b00e24e1a7501c6d
66d41e0c3175b1a4770de642b284f1bf4b021552
13304 F20101123_AABGLN guo_h_Page_031.QC.jpg
d4e9483173ff14d61faae45d2247b20d
c7f663da966139c4d1cb3bbd9cbccb20c2cccd0a
5989 F20101123_AABGMD guo_h_Page_043thm.jpg
6f06b3ffcb8e8ae45fcf56db87ab29cf
9c41c7541562c19136cce41fdef2e074fe0843ad
19446 F20101123_AABGLO guo_h_Page_032.QC.jpg
22e8c4d9bb1d2215ba06d7f5eef40df9
ca6862fba13ba684f58c9fa6d341dc699b414da5
4817 F20101123_AABGME guo_h_Page_044thm.jpg
7203cf53520e8041236492a3635acd2e
1477c3672afa9f26448e07154a693ed226c85860
13997 F20101123_AABGLP guo_h_Page_033.QC.jpg
35ec70424fd7e2d8a5aac59c46899161
346ca190ccaebb923042b998a075bed4be72f270
5987 F20101123_AABGMF guo_h_Page_045thm.jpg
a877fc33384a5fc553cfba05da7af36c
005e4851c6a301f32e44d3fefd8a7c77808ff30e
13121 F20101123_AABGLQ guo_h_Page_034.QC.jpg
dd7dbdde37ecea273d9fee5ecc308488
d85230d18e37cceb1fa8e5f12ef9295d570a1413
16071 F20101123_AABGMG guo_h_Page_046.QC.jpg
3ef8b2b08086ece21d97c6cd236662c7
1358687ad835c6411fd8b4443bc854a3dc18ebb1
4066 F20101123_AABGLR guo_h_Page_034thm.jpg
85a821ef4e71cf47716f979f0d72ca4c
c76037c79eaa15cccf9bf7bd7a3994a3f880fcc5
4774 F20101123_AABGMH guo_h_Page_046thm.jpg
52d073f8fe68e30d65673f99f10410eb
6e019fdc949ac97268529776ecbad7c4f212f36e
19171 F20101123_AABGLS guo_h_Page_035.QC.jpg
9daa1b30f54bb6548c506e9ea0228164
7457e7374b9cb8cc857c4f8f0fdcb2cdcc3fa597
19956 F20101123_AABGMI guo_h_Page_047.QC.jpg
dfa07d29c0c4f13495976cd786052c48
895d24c8851218039ab439a520919a43e6c04a1c
F20101123_AABGLT guo_h_Page_036thm.jpg
3ff1568cc69169db20f64bc685800824
3cd95f62237602ac79acc26ebc351afd084e64f7
12853 F20101123_AABGMJ guo_h_Page_048.QC.jpg
9c1f3f5d49384e5381fac039be138fdf
e0b64ae4a5bb633f91abd5418fb2f76cc6fd976d
11310 F20101123_AABGLU guo_h_Page_037.QC.jpg
6edaabde668d8ebdf18e101759c10d56
93ab12baae414fd09ac058687f4916dc330c2983
3818 F20101123_AABGMK guo_h_Page_048thm.jpg
4b044d455a2eca5bb1b60ba33bcaa571
0aa5c48eac287790cd3c15c285ba3db694f33e29
20332 F20101123_AABGLV guo_h_Page_038.QC.jpg
0fc9d7f9ce2e6848a577d6c67b7263df
c85fdb737671f8bed9162a9d4ee575266e1bf3d1
10962 F20101123_AABGML guo_h_Page_049.QC.jpg
f0aeb8163f65cc133fc438261e13c04d
39c48c584fb94e01e0d954602a523c0ffd3b6835
5580 F20101123_AABGLW guo_h_Page_038thm.jpg
8a26b2ccde43b67f1765dffd13cee2b5
22f01f1211d0b42b486d6ca6ab64bedd32178e89
21021 F20101123_AABGNA guo_h_Page_060.QC.jpg
01d6dfed999e51cf524106a513595bd5
fcbea2edf4f738c6eec168a060f95e84f613c20a
22837 F20101123_AABGLX guo_h_Page_039.QC.jpg
5462b4667fc77dc9d39de94b8750ae4e
41149236cb84c2bfb1bc477fe027ade14731627c
5695 F20101123_AABGNB guo_h_Page_060thm.jpg
4830bbafae4d443e5e8dce594ba96b34
ee360b8d91a8fc35aee3a34536cf146f1c959552
3385 F20101123_AABGMM guo_h_Page_049thm.jpg
70870e7b56737701a39b8313c6ffea00
25429ff5d91c51ddbc7bd5bdd03d0d93af370e8a
25603 F20101123_AABGLY guo_h_Page_040.QC.jpg
50f3388a2b5e226b15d33e4f7739fb8d
b194c1dffb95ee41e8974772aef682996d5cab9d
21983 F20101123_AABGNC guo_h_Page_061.QC.jpg
038047f2ca3c571d27a5354e7bc77001
1127005e073629fa088fd884cd329eb4c58f0a80
11768 F20101123_AABGMN guo_h_Page_051.QC.jpg
91dff8f4ac12bcd3b8d1f633905ca701
15fc6c63d60142871139291d4bfd4fce6483cea2
6623 F20101123_AABGLZ guo_h_Page_040thm.jpg
85fb93f00d0a2ba37e21528ab0d2e311
7f4829a44bc09c12a048bc36a247700af6f3f907
5758 F20101123_AABGND guo_h_Page_061thm.jpg
1500e62409ba893bbf8cdd59d7991c7e
167d1ec5797ac2894a45f2ef470fcda82f1d3803
4900 F20101123_AABGMO guo_h_Page_051thm.jpg
6906e8c4279c28182e5050d86ddcb1be
0206762917d04972ee0f70a4d41643ce441471f7
11941 F20101123_AABGNE guo_h_Page_062.QC.jpg
d4a31eb4dd03d2ad5a1774c0c322bc5f
4f292391aac7383fbafcb854ea5dabd243f8be99
15835 F20101123_AABGMP guo_h_Page_052.QC.jpg
ed746cde6719641c2d9a5499f233a9bb
ed462786a1edb0c2bbb1275e8bfcd9be367899ff
17302 F20101123_AABGNF guo_h_Page_063.QC.jpg
3f9a7fe18d5359764cc0914d311f8644
7b4dd6cc70823f1b61dfa0c4fafa75aef4177753
4009 F20101123_AABGMQ guo_h_Page_052thm.jpg
1988c80f7cc79c758a539be19bde5137
f2ccc2bc5958f33ae343659b54f2fd4298496853
5153 F20101123_AABGNG guo_h_Page_063thm.jpg
79ba53ffde26e28205ea4b6158ffd83e
3d43795a541e153875dc938a9b1bc56f54cefde0
5955 F20101123_AABGMR guo_h_Page_053thm.jpg
d8c324e29e65fd294c1eaf51691e1bda
8b2bc9929bd79e54dd1f021089bf33d4033e63fa
4927 F20101123_AABGNH guo_h_Page_064thm.jpg
3ce391c1b76b5fb75c489148201a7c32
97feb543fd83507f908f30795a75262033f711c3
5484 F20101123_AABGMS guo_h_Page_054thm.jpg
944f540caa87adf1e83ab68b0ba592cc
c313decdf6169b4b65c85eb473c7106091fc5902
18155 F20101123_AABGNI guo_h_Page_065.QC.jpg
44691fe22bab67c5bf02f59bd324f373
1563e0af59b71d5530ab9c93768c08411ef3746c
4387 F20101123_AABGMT guo_h_Page_055thm.jpg
ef89436305e29e6493043cc5a12c119b
41ba5a88aa4d575a0ac2646977bdd5c798951ce6
4895 F20101123_AABGNJ guo_h_Page_065thm.jpg
922b31dd22fa41d14621e070c7c7d697
0174eea455501db623912158e6f8317103615e51
17563 F20101123_AABGMU guo_h_Page_056.QC.jpg
b12208b264686ef42f99c7f0f9657b57
8a78917e81f9da5b7f54688a1b989c4c0a4be8db
5696 F20101123_AABGNK guo_h_Page_066thm.jpg
fa51dff003f5300b909865c7a222dd7f
a5756cb38c5dfe070c1abf712d1858a62debdb51
25380 F20101123_AABGMV guo_h_Page_057.QC.jpg
26f2f933126710300746994bb239fe9c
b2bb5cfe50cea0bfd4502f47530d815869cf025f
16197 F20101123_AABGNL guo_h_Page_067.QC.jpg
c4b82d4c8ab3122271ed27fc12d7d78b
83a661d358f555428ad53d333e29135ae5ad753f
6420 F20101123_AABGMW guo_h_Page_057thm.jpg
5ff31f0f39af87a6b88c2e68113ea2f0
0c6b740e606aa7770765a4e2601b28465a7f7e4a
5095 F20101123_AABGNM guo_h_Page_067thm.jpg
cb5192665c4f79d572d61a9cf354c6d4
56ee285a6bc9bed45caeb7b92081a68858d68e86
13031 F20101123_AABGMX guo_h_Page_058.QC.jpg
9dc077f65b5584a38dd6df777c38160a
94aaa80ed19b066f3cfc5c30ce289751ba9f9685
6511 F20101123_AABGOA guo_h_Page_076.QC.jpg
089f1e18269c72b13b0bab001dab151e
089010fddff92fddc4884255118f6336884757e1
3813 F20101123_AABGMY guo_h_Page_058thm.jpg
808d6992a5b4dd95c5afe809096b8b7f
7384333ff32b6b9155941e9584baf117e13774d9
2279 F20101123_AABGOB guo_h_Page_076thm.jpg
75f34f3aac343190a5bc2710fc83bd85
02c9d848ce5d4c97c5dddeb7db29df5615d46911
13467 F20101123_AABGNN guo_h_Page_068.QC.jpg
5458b1dd299d293128897eeebae8a706
475ab8965516d556945858e3530334f06b3df157
5952 F20101123_AABGMZ guo_h_Page_059thm.jpg
f096fbf8b1795760999a8e5a426ddbe4
055768812c771132f5a0632cc77be01715cdcf15
28004 F20101123_AABFLA guo_h_Page_080.jpg
d26bbea47d0a5ca1a83055908f1a3f08
be1e425e7d0470fab0b232c5f0a749b7746bed9a
7675 F20101123_AABGOC guo_h_Page_077.QC.jpg
030236d62b18d013ab0ba0e3557d2010
3a397461ff4287881dd7499dc7eaf0d54e73e575
4125 F20101123_AABGNO guo_h_Page_068thm.jpg
57d7cd4ed2e03502b81fe993d26cc893
6aa815bf3e277c089929cc014e713b0e065fbcc4
54136 F20101123_AABFLB guo_h_Page_094.pro
8ad6a77aab82df244cd1bd12e8d1f11a
17d71048a904cc3c58b33788c6e8eecb948b9ea1
2510 F20101123_AABGOD guo_h_Page_077thm.jpg
9bec6545aa390ba1bfcecb9201f01821
87556fb9ac85b55211124cb9e08500b32020e09d
15762 F20101123_AABGNP guo_h_Page_069.QC.jpg
722dc372083e8e5efd2867d23eaff847
94b9b516a5b8bac58268cb17dedfb47c8ae7b134
2231 F20101123_AABFLC guo_h_Page_014.txt
9c4d8f668770b17dd528a7c67195552a
344eb1e0988f4244051622683546375352f3ada3
6981 F20101123_AABGOE guo_h_Page_078.QC.jpg
7986d9925df61da0d0644a94a11fc2e6
3704fef49f7b068ac00edb9e9ec5bddbd03a9793
3939 F20101123_AABGNQ guo_h_Page_069thm.jpg
bc87a48a7708987721f508f0edda1cb7
a82a6ab299214fb69d294f25e89c34fb1473e466
F20101123_AABFLD guo_h_Page_012.tif
d9ec0ed2a6aec615de264b6dfa32ff71
14d7c87eadeda29fdb8a06871fb2ba203dd9a156
2380 F20101123_AABGOF guo_h_Page_078thm.jpg
358842e9a525b1b49c81dd4365bb4bf1
45b259d63137dc80966a7275d73fe249457ef12c
20456 F20101123_AABGNR guo_h_Page_070.QC.jpg
aab151578a0ab936ba689429f9c8487f
5c43e29b1cee3f6727aa670025eeba98a7957a68
36279 F20101123_AABFLE guo_h_Page_085.pro
7834939307b211b5f8e2a1dd6389cb59
128d23a00c5a635a69d3fcb0b28cdbe2bc4729be
7360 F20101123_AABGOG guo_h_Page_079.QC.jpg
e1e821816dc6d036422a01290bb83a5f
704be05bef773c2e1006affbecc44022884011a0
5575 F20101123_AABGNS guo_h_Page_070thm.jpg
f30c4d91461a78d21d90ac2fe5bdf256
d49ade7a6dcfedf61f21d603fa3ce20aa4bf5d7e
21026 F20101123_AABFLF guo_h_Page_054.QC.jpg
ff4335a9e53e4310a6deafb49c0dd34d
12f365eca529735d701a65a4bb2d09465709b2cb
2378 F20101123_AABGOH guo_h_Page_079thm.jpg
d473832b3ebea01858bd797ad13a9af1
a3d51cef6b5d78e4919d7111148071c3ce2af575
27078 F20101123_AABGNT guo_h_Page_071.QC.jpg
c12b72f63517acf4d1d7e3f868b45bdd
6365506bc802f41fb077320c250b8839bb8a7131
32405 F20101123_AABFLG guo_h_Page_046.pro
6d5aa84abcf5ac82f314688d710c0af7
614bb75c0af465f47d40c799fc5270e50285b31f
8008 F20101123_AABGOI guo_h_Page_080.QC.jpg
3acaa4c462a2eb557c123e84d8091c45
206e87216ba47d95383bd3af78575f0ce4140df0
6668 F20101123_AABGNU guo_h_Page_071thm.jpg
673cf5c2cfa0f99e54c1e2cc09a6b4d3
2208df23cb3f5dac0799f2a7e22d9873c8e90a0d
56672 F20101123_AABFLH guo_h_Page_027.jpg
74e80bdf105fe535b23a3884267bac33
d4df019ead81a3418361b16d2ad49b8d8f94b715
2386 F20101123_AABGOJ guo_h_Page_080thm.jpg
4bff83564df224c5dd26ed2b807dbfd2
47743c2b784780505e053d03dabdcd9c7106e1c0
23894 F20101123_AABGNV guo_h_Page_072.QC.jpg
9381fc18a163e62d709a2a841ec63209
2e855e090b24f09dd23133d186de8ace9a61372b
11104 F20101123_AABFLI guo_h_Page_080.pro
21666c1aa015376c1e04295cc0ad1838
ffb35a389f7417fa4d6fbf2e176615bd4fafefd1
17404 F20101123_AABGOK guo_h_Page_081.QC.jpg
c267e7275c180119ae784bf6ad97d1ba
0ff1b35926cfc410978d1387a4d5252183a0bf47
23187 F20101123_AABGNW guo_h_Page_073.QC.jpg
9b97ba92750c0c8b33ff80c9efba3d13
fc9778275ce749c27bdd2d8f6d6256e7df95f73a
1861 F20101123_AABFLJ guo_h_Page_011.txt
76aaa805d90ff06121d422d1bec9637f
7703b1f27e2731b15081d898cc9195a2d0ec30a2
4429 F20101123_AABGOL guo_h_Page_081thm.jpg
6ac75fa428261479e601ca43fb015645
196622158bc47cda2ede4ae1c3dbbba8fde1b7a5
3832 F20101123_AABFKU guo_h_Page_082.txt
932b3d3bf3ad3edba901f4d32bb92d91
f98d97ae42654d2aea33158995460c2831efe987
2917 F20101123_AABGPA guo_h_Page_092thm.jpg
e885a8d4cc4f4d340f1aae7969814b89
8ff16fb92a5d482cf605cfe6ca1a1023df24d452
6110 F20101123_AABGNX guo_h_Page_073thm.jpg
1d91047a973e6f2b4d3672a12a39acd1
02e1ddd167936ae4a7a7ca08a8ad2dc307d04225
827119 F20101123_AABFLK guo_h_Page_064.jp2
ae3f35478a870546943f1a4b58b2971f
8423a3a8607f60006293969578d9ddab81b0612b
6141 F20101123_AABGOM guo_h_Page_082thm.jpg
05bbb1811577281aec703636f614107a
6c9d13747d7b0d93ec96d675a35e057b754edf8e
4945 F20101123_AABFKV guo_h_Page_049.pro
156a89dd87e58aa07abd5a2290f6010a
7cae5a0c0daf92e284b8cc4db1e79fe928419f27
16392 F20101123_AABGPB guo_h_Page_093.QC.jpg
b94ff390f890664dbffa8c80d5d168e7
b0ae8f8ca39eed146043918f25aac42171ec3326
20108 F20101123_AABGNY guo_h_Page_074.QC.jpg
e401f7dab4923b278d85dad8151253bd
455e91c71e9108df2c82f596a8bd3ce729db144a
17915 F20101123_AABFLL guo_h_Page_086.QC.jpg
ef5f0596f11ae63b73dc29ebaaf27df1
e77ea2aae579c0775840b820bf7e149b02005e9e
18868 F20101123_AABGON guo_h_Page_084.QC.jpg
46263e1d40c0afb0f5d65dc332eba592
13e1f52515e1b5f3972369935abcbd0bfdd10a7f
4030 F20101123_AABFKW guo_h_Page_031thm.jpg
285249b45c8f9a067d7ac34935d46208
331f8900f65bbe3e03a1f67e9eef851bfe13633e
4813 F20101123_AABGPC guo_h_Page_093thm.jpg
82fa7de1d9b141f51a18932fd5771fcf
adcb5056979b0d601ab824c509190d97459287e3
5510 F20101123_AABGNZ guo_h_Page_074thm.jpg
6450043161a2a5ed7c0a04331ff24beb
88fc1202f2f3aa789c67a4927bb6f90aa7379b66
F20101123_AABFMA guo_h_Page_004.tif
793344922376658a741eea5c0f0fb7a9
e93a8d4039dd7dd03eaeabe286d284874827a05f
F20101123_AABFKX guo_h_Page_074.tif
1d43dcf17dd5e2ba37d17464290ebc09
43f7c4ba40a3153d36d48be23dcc364e9c429c3e
6629 F20101123_AABGPD guo_h_Page_094thm.jpg
c79d24fdf02d976fae7202d78dc29c73
2258d65a8f7588382ae114dbbcb17bd29b40b48b
29715 F20101123_AABFMB guo_h_Page_093.pro
8f8ae1aa8bed0e1f4d034dfac16627ec
c755b717fba18e14ed548db144259dd991c4fede
19480 F20101123_AABFLM guo_h_Page_103.QC.jpg
d71c23d8e6e07063b705cd171f30e29c
c09ece83224e68c437f31845793a44309f4893fa
5511 F20101123_AABGOO guo_h_Page_084thm.jpg
0da9cce1c5594daca3c76efaddd303d3
f060f23911070b8002a1145b30e4205d53faf613
213180 F20101123_AABFKY guo_h_Page_116.jp2
ac64f12540a52fbd7abfde8293df21db
98a1b745fb069baf59a8655f9a5a14d218c7827f
18065 F20101123_AABGPE guo_h_Page_095.QC.jpg
0cfa45b0c21e51324966cfc6b7f8bee5
7eab6fbff6225f28380ccb27aaaf3d740f30f798
F20101123_AABFMC guo_h_Page_017thm.jpg
20abb4b416d1f4a04515bb8c58ae2f0c
068532df81fd11da2c25d38c5e1b1750758eb367
1289 F20101123_AABFLN guo_h_Page_034.txt
bff97ef5331b319374377c14fe7bc767
70300bbdbce5d9f1b6e99135b6c95c76c43b0844
18309 F20101123_AABGOP guo_h_Page_085.QC.jpg
dd58d10d969f20e6bb1145410bbc49fb
60788ddd7aae4102342f3f0289d3a81aad791a66
F20101123_AABFKZ guo_h_Page_040.tif
963bacbc03cbe68fb38e2ce592fb5b8d
7fd826229ff44ec6979de21313be6b6da77b4b17
20917 F20101123_AABGPF guo_h_Page_096.QC.jpg
4f273ff022cd5f878555129690b197f1
80eb277c7c7fdc883f11e48502cdb2fee7593b66
21006 F20101123_AABFMD guo_h_Page_076.jpg
eafb82270d2648523b448d38c5c8e87d
9413c3aaff12169b84778f9cf2e656c2c1dc8854
818731 F20101123_AABFLO guo_h_Page_007.jp2
e360fd67ef4d0f4ae70f5715422a76a8
2adca6a5c1084c0be2d0aa2ee407117d302a80bd
5104 F20101123_AABGOQ guo_h_Page_085thm.jpg
9718269634becbdbaec52a9a4c4d74a0
75120a887a7dcfec68ec07ad11cf280fc1cb4254
5375 F20101123_AABGPG guo_h_Page_096thm.jpg
b291abd2f39d8630d1a3a98954864d4b
06deddfc1c9c6fbdcfa5b2df1e0479ffe619fe8a
6007 F20101123_AABFME guo_h_Page_024.pro
d16caafc9d362337f387bb52fd25fe95
35f299a1bd9b49a0bad7a3396bbb8ca3a46dc9db
2435 F20101123_AABFLP guo_h_Page_105thm.jpg
16b26407c9729003535c7710543f595b
a863fc0e6c070d71c1666331d6a9079cd39beb33
4917 F20101123_AABGOR guo_h_Page_086thm.jpg
77ee100ccc6bc97401616b818622fdd0
e1fc641684ee8caf05d61281a5d3b856de6b457d
8656 F20101123_AABGPH guo_h_Page_097.QC.jpg
3b6ee0f89edf44c1f03f13aa95dfec06
e801db73fcf270c93c4384284098ffa69ad3946b
F20101123_AABFMF guo_h_Page_023.tif
5eb05d172abdc3e4b57487e5964caab8
820d271bf524091c5dee55aeabca5d47c5c02366
37148 F20101123_AABFLQ guo_h_Page_063.pro
9c48d8145b5a8bfe8e2749354aa1332c
584dc4fb45e354cb474db76f30181684c44d85dd
22725 F20101123_AABGOS guo_h_Page_087.QC.jpg
b39a0d212241c82597ddd2fadc797da1
26c3617f520f74d07833ba637b3eba928eb294f7
3168 F20101123_AABGPI guo_h_Page_097thm.jpg
c4ce0ebdc255631e5222560401b53364
2fe88e3986bdd8cf287b0e363a0493a32c487f98
25124 F20101123_AABFMG guo_h_Page_048.pro
290f1ac78f00366e04100feb106568c5
12ad7066c4d663cf2adcfed5ef7e6fbcc03efc9e
168 F20101123_AABFLR guo_h_Page_077.txt
1bb348b68d38de5150b1dbc2d9d5d1d4
65d4ae2a4cc7c386188dd645bf119a9dec0cda0f
22392 F20101123_AABGOT guo_h_Page_088.QC.jpg
b8ae45c6ccd633d02d40af3bf65471fa
710f238963fc34e4ea9040473327d100d7a8351e
5847 F20101123_AABGPJ guo_h_Page_098.QC.jpg
3f67f4c5b23dfa19e9eb96469486487c
c588f1394f0b9deeb9d51a438b56d5f43b2396b0
2780 F20101123_AABFMH guo_h_Page_104.pro
d87c6ab11e7563514a7ddfd477597298
abca5349f1a5db466ee2e8da222858a1f7a7389b
F20101123_AABFLS guo_h_Page_103.tif
a7b7b3ee748484be6b9916f1ebb71366
663eb64402f5dc5055192f21ff1ea07c79bafc3f
6191 F20101123_AABGOU guo_h_Page_088thm.jpg
2030679985ce3964423757cdfc4b5f92
5e8be87d8d70557925e38d07ab52a21525f97c0c
2100 F20101123_AABGPK guo_h_Page_098thm.jpg
981236537ff2e02d628e845544d484b4
2116dcd9a0e09d28c1b631c1d6cb1ea4b9c0f906
20947 F20101123_AABFMI guo_h_Page_053.QC.jpg
fe48bafbe91d5a8b79c2407f88ae7aaa
affd4fb9ec72da47155b3b5ae886d1e65ce0e041
346 F20101123_AABFLT guo_h_Page_076.txt
a03b679541e3965feb40cb82bf93b5d3
00258f290c406e821c15ba994323ae191afaf49b
22004 F20101123_AABGOV guo_h_Page_089.QC.jpg
e9753d85c93036592d825d45af301861
b4812181f7aef6464aeed7c7b8c299b757877d71
2038 F20101123_AABGPL guo_h_Page_099thm.jpg
d132d94105b9ecd79e0a6245e8d83d1d
ba19893baa690ffb3d6d1beb5d89ae2263926606
24762 F20101123_AABFMJ guo_h_Page_105.jpg
8ed25ba7e9229c2de7bc1d8c5b2604e0
8ff072b211fcdb6cc22b4f141a15b15b9e37ef9e
F20101123_AABFLU guo_h_Page_108.tif
157e7ba01c2a8e032bd2a4b3ec1784dc
f29e52bad8f940abe96b350b3f2405b7911572ba
24264 F20101123_AABGOW guo_h_Page_090.QC.jpg
e679e8942246ffdafddad5cef46f8df1
09ac57e04d7d7c7f893c5a3019847174b0a31dc3
15256 F20101123_AABGQA guo_h_Page_110.QC.jpg
051017bd93c86e91a70592e193434412
ee8b470bae6bf66b9bcdfa6fd912ebcb2cdce899
16178 F20101123_AABGPM guo_h_Page_100.QC.jpg
519cc8779f6895fe8881a0f76b916313
d950b64defacccedb571af3dd93f505cd86110a4
50481 F20101123_AABFMK guo_h_Page_072.pro
78f09fa36554d6079ae20c367907c91b
b02ccdc615478b29a067ffd2799249af86c8eb03
1859 F20101123_AABFLV guo_h_Page_019.txt
f0a0cb473234761dafc7d9d29cb6d5e0
30cb171e53f81bef172652bf735ca6f52f14274e
6297 F20101123_AABGOX guo_h_Page_090thm.jpg
4db36060a8771be22d2735fabfdc5bd5
b6c3627264a798c636a87660fbac203e563cec06
4212 F20101123_AABGQB guo_h_Page_110thm.jpg
398de3635d75175b1e13fc0948bf49b6
ad2343c9768f7df8c70df27bf9bcbab2acaa2c0f
4677 F20101123_AABGPN guo_h_Page_100thm.jpg
49e95ef79cfced837eb14420f8d30bab
45a5f0fa83f9f18f93699def66f5b6f51cdbf58c
1595 F20101123_AABFML guo_h_Page_091.txt
5d99e94576f31702941b0c8b27ba4035
40ffe0af9d8caf3ba93b2a64bc1ad18d0902d598
21129 F20101123_AABFLW guo_h_Page_117.jpg
e6a86ce716ee0af5428e9439c8dd3ce0
1b99aaae759afa4dbf57f0dbe08cb833b10d1910
19661 F20101123_AABGOY guo_h_Page_091.QC.jpg
4be857edbdf50e4cc7e4bc44a1851eff
8d43d6c7848590fed1ed6804cc3323133f10dadf
21901 F20101123_AABGQC guo_h_Page_111.QC.jpg
5e6e35a7890f9198dbc3b28380561239
f299ddda7ed41ac88dc0306de040d9391ce0d12e
F20101123_AABGPO guo_h_Page_101thm.jpg
5a0a64bfe116a8fc3b20e4a9af9569d8
dc715f655dcf2da322b6024a04cf5863cf362889
7553 F20101123_AABFMM guo_h_Page_101.QC.jpg
604603c7d85744420b5e778cb342524c
ffa43ca72d91b2c825b7f9565deb03e5d1f4e004
280 F20101123_AABFLX guo_h_Page_020.txt
649237b5d732dd061c25ca78d74b5571
0808740f8c765b2d77ad45086bd9fe733beb2d16
5367 F20101123_AABGOZ guo_h_Page_091thm.jpg
b17a3ccb207e16f12680a18327d922d3
fc6f1577fff78a5e41f472e6fac4d9b37efbe272
F20101123_AABFNA guo_h_Page_070.tif
6500a29ab1ffd892cd4ae4210df014e8
b09f86237f6686dd20801841238db1ce8cfaa4c3
5960 F20101123_AABGQD guo_h_Page_111thm.jpg
1fc7d972b6d0700a21b0ee9ae543ab26
512d0b8734f86e2ef004b716cb549092fe10e801
1854 F20101123_AABFLY guo_h_Page_042.txt
72a58b9540eaf4e1af7ba3ef097c423f
08d2c3e3678ad2dfcb38b9cc6e2bd40f49dd850d
56714 F20101123_AABFNB guo_h_Page_014.pro
eb8c461bac9c9eb2e3ca0801c232ad79
6b3dc7cc22f5e2867113b311dc679ac6501dbcd5
6728 F20101123_AABGQE guo_h_Page_112thm.jpg
ecad5c276847b51b035c14b4dd119d5d
d41ce0beaaacd3508b51d2eadd12c0022f33131e
6643 F20101123_AABGPP guo_h_Page_102.QC.jpg
26688b52aaa2ad57d9565774408678e5
2d00756c39ea35a0fae21d43956264bcced06dc0
989 F20101123_AABFMN guo_h_Page_005.txt
59a32fda0e43196a63e5f7e084556503
a9e3f560d51cc225cb70fa2b88d900e1541a4de0
F20101123_AABFLZ guo_h_Page_083.tif
06ce3a7dd9d16f684e2f89c4fa167733
31e9bb2533a14bdb47a591a53374e72459479d99
694442 F20101123_AABFNC guo_h_Page_093.jp2
24428b1eca6e761948e2b9f65a111567
25ee1bff79e1db7dd67510549357075355db47ff
25643 F20101123_AABGQF guo_h_Page_113.QC.jpg
5e2bd213b221716d53f63aa737d2ead4
70ba5d1cf6bb5748d770d06201ee31d539a55f83
2329 F20101123_AABGPQ guo_h_Page_102thm.jpg
56930f22a4723b96e65c27f0b56041d4
eb782f2b89a3c4e62eaa77317ebb0a3581ed9a00
1762 F20101123_AABFMO guo_h_Page_004.txt
b29e2012d726ea14bf795fa97f926eb2
534596f57657c538797afb2ff564df7f645deb46
22158 F20101123_AABFND guo_h_Page_102.jpg
f348e17245011b2019f0c951b0bbaaa8
5ad29cda8551d8195a6070bc4e7d1f3917f6050d
6665 F20101123_AABGQG guo_h_Page_113thm.jpg
5dac50288a9d16a43cdf54d789739b10
e33adaea8ebb2f48bffbb96c8a215cf9c615999b
5307 F20101123_AABGPR guo_h_Page_103thm.jpg
89ab5f838ba915270ad37d049902458d
4b6ccfbc3ba1335da74e7fcabbf47045ad0b5c73
F20101123_AABFMP guo_h_Page_042.tif
9199c1fb58eb2f0288046d1d3178b361
c233b0a2215e7b48b6fc472fd3210c36f4617ea8
1051965 F20101123_AABFNE guo_h_Page_010.jp2
6ec7ed896dc836274bfc1627ff42426d
eeec922e11f90887aac980ee58a0e4b8a297e983
25087 F20101123_AABGQH guo_h_Page_114.QC.jpg
edc72fa8ee8741ee932db53a04d3f400
549947f74718e23dc193b0f18f06a3a21747825f
2626 F20101123_AABGPS guo_h_Page_104thm.jpg
a0a06b0d87f9ad604983088b2ea6343c
6763782a37f0f2951724cf7dbbdf7dbc9404e696
F20101123_AABFMQ guo_h_Page_021.tif
c8ca3b1364f079d6ecbb04de2115cf68
575e8c2f138578ea352b077c84cd6eb3a3452faa
478760 F20101123_AABFNF guo_h_Page_104.jp2
3322f14b5ef3df418a69031cf044da24
570febfe5f0074e029a740477b830c2722e7e523
6598 F20101123_AABGQI guo_h_Page_114thm.jpg
e4e0852f2323d8279d233d2f4c34fe0f
1da366293a318b5bfee735d47ed8746708308f6f
7514 F20101123_AABGPT guo_h_Page_105.QC.jpg
da6c7515094ee12de1af89a12b8a6756
53247700b1632b9b035995214c35b5b9075f81aa
1007534 F20101123_AABFMR guo_h_Page_059.jp2
58d8ba8e8ce97e39173ff67bb2e786e8
bc2a6af694ff85b3e846192aec8f66a6458ff87f
54771 F20101123_AABFNG guo_h_Page_022.pro
482cfd9773b26327c12ebf085ec2a8f6
90577bded152c12c841a5576dead9ded53f9a167
2079 F20101123_AABGQJ guo_h_Page_116thm.jpg
edc59d2c6aeb314326df167cada8fc05
1fbe8fd4af8cf09e2e51ba6253bf6e65a45ad14d
2885 F20101123_AABGPU guo_h_Page_106thm.jpg
90f819cb43c3bf78fddf7fd58133759b
966a0bc85e6dfa4decf0c3bc98497c4ebc3dbf41
F20101123_AABFMS guo_h_Page_062.tif
6999628e0f4ff2c16c75cd1843bc2ccd
b9cf04c5968759546297d8e10bf6f0e693b1f29e
32055 F20101123_AABFNH guo_h_Page_056.pro
0c21db7a98a97316a8b56800543e5eb4
4b305722ff8870372877cf693495b16d44f19471
F20101123_AABGQK guo_h_Page_117.QC.jpg
ac5c74ad5b440cf9a5826d5788db7a3e
02ed9a6cea910457b9b2e03b8f41f798ebe13d27
2704 F20101123_AABGPV guo_h_Page_107thm.jpg
0188a941d25ffe1ac1b4c98b028de9b1
42bb7cfc3bcbe2374b3d981d1f9906aade8c1000
24953 F20101123_AABFMT guo_h_Page_094.QC.jpg
f97c227d634ff2c11b4e9260b0082ccc
246a7ca30dc078093ebd5d663caccf12f2d9670a
F20101123_AABFNI guo_h_Page_046.tif
308fb6da0514bb7ce22e5f3443e15bee
21b37799d4c12b6b820ad663d391ba6a9bae413d
134018 F20101123_AABGQL UFE0011645_00001.mets
ec92839ee425286da7475c2c007d71b2
ed90319608ba300cd3c2de66c70ca21e1b68df92
6259 F20101123_AABGPW guo_h_Page_108.QC.jpg
0745107906849c8494f7852f456b8f25
ccf86d390cb0df88f73213fa5b4ad1ff70a92e23
6176 F20101123_AABFMU guo_h_Page_099.QC.jpg
34ce01b45b5c0e9d8ce813efb119d8d7
4e35e83d0624266a9aa65d0e3a2cd7883e797297
20476 F20101123_AABFNJ guo_h_Page_042.QC.jpg
3153660ec5412dcf1a578d3370bcfdbc
52a350eb48244c5b2da5ecb2f27faa3488ed1f09
2288 F20101123_AABGPX guo_h_Page_108thm.jpg
6e8c0b3b0364f943864a1228b5cd045b
8bd194d224bb0b3ae668d10450be1cd54b91b5d4
49718 F20101123_AABFMV guo_h_Page_028.pro
eac9f0f8b23ef8be36b722893ddbe367
1c20050feb948fb1f1c3b26ded282a3d56119f61
640 F20101123_AABFNK guo_h_Page_021.txt
558d2e4d90842af487004930395a146a
cec74ef4cea142ff13cb5abb22adde71f542c445
22928 F20101123_AABGPY guo_h_Page_109.QC.jpg
0d41884f9d6fffca62d950e11022dffb
1e4d2f2e43db17f9fd276ba5092eff34323583f0
3718 F20101123_AABFMW guo_h_Page_037thm.jpg
c2564ad9983008e7ea0b3c774336169c
fe3ea41e76328d92a61ad02a6fa92d174cee5b54
2096 F20101123_AABFNL guo_h_Page_059.txt
bb0cace293dff1d12432a30cd8dae6fa
44b01683eaf9e79fa5cb548c00f98a890bd3a378
5807 F20101123_AABGPZ guo_h_Page_109thm.jpg
98c152321fa86c086b8177a4bbc86772
9d1757ecb96719c8f83002175efc7051a4264db4
38639 F20101123_AABFMX guo_h_Page_091.pro
a0410e04f13b3379a0555a3320d5afd2
0438fd5ea836ee301037223e5ae4352f820b273c
944212 F20101123_AABFOA guo_h_Page_060.jp2
55ee6e6f446af1cfbd76f357c3e7d673
54b76496c9ef1a4b7ecbc5cb60412ed04518e20d
5912 F20101123_AABFNM guo_h_Page_089thm.jpg
a2dcc5d924749e87f348030d122e7be8
68db439b86e82b1e9ae547d10ab0523792233c95
61497 F20101123_AABFMY guo_h_Page_112.pro
80ce6c3ec04f24da39fd85f68e79d300
a4f2e2a32cdf313a9cb712e15e76c6a5ebca27da
78142 F20101123_AABFOB guo_h_Page_111.jpg
5d14c94c9bc6bab8729408cf1590979d
59fd1b79ca1f737c979389cd9fab745a854fdd89
29026 F20101123_AABFNN guo_h_Page_069.pro
b925c73cb2e5ee51968f815e38712339
6751f7bc87e952c5c6421f921eaf958283ebcf6a
8543 F20101123_AABFMZ guo_h_Page_092.QC.jpg
caa8cd22a5033b313ffc7a6c18e598a3
3090e6f07d7c414d996121b98e8c3208ace7d063
F20101123_AABFOC guo_h_Page_014.tif
b44019fca7322204397ede577f898fdf
a39e6f9ecf94d0bb471a376bce74f270ee6b93bf
444 F20101123_AABFOD guo_h_Page_117.txt
12d58b808a3be5fe86f2d27d106373b1
ae3d0ca862eb9042d1fc146633e2a03d4aa90d83
5146 F20101123_AABFNO guo_h_Page_095thm.jpg
f3e91e54fb5a483a23cfda05d98d2b86
6dfb1ab749b8d9f48b4a382182d93a864c4c9826
16367 F20101123_AABFOE guo_h_Page_027.QC.jpg
043958f6f68bfc9ab6e16739b8db3366
2a61cd1eb6120b75ac83bdec2d6e847ec12a2a18
1200 F20101123_AABFNP guo_h_Page_110.txt
cabb13c31203a88b5ef1a0cbc72e3c3c
55d8e5ff4af625ea7e3fc0da84f9a7109934fd0e
F20101123_AABFOF guo_h_Page_078.tif
1bb5926bf4dc6b1742cf7ef5f8ec6ca3
25113eef3fd385b929ce00ddf20ce379e34db16a
64269 F20101123_AABFNQ guo_h_Page_038.jpg
b073755b9fda6bee58073bff683fb70e
e188d3792b5d9a5ecc24647e105e782ad4bfe6cc
1051938 F20101123_AABFOG guo_h_Page_057.jp2
454c8f1922ffba039ec3398133b61246
3d1e06989fc33c37221ad7ab30a069a42be4ddbf
5708 F20101123_AABFNR guo_h_Page_052.pro
e290a1c7c5bbabe397d86905c2640ced
79a74aa9eeea1ad6d83faf9ac6df791bdbbca031
1051975 F20101123_AABFOH guo_h_Page_012.jp2
08f53f79b135cafb2c473753bc831494
2f9c0fb14eb425264b646d51c9c832da9fce6203
5951 F20101123_AABFNS guo_h_Page_083thm.jpg
fd299b49c27ddd856646102544d3813f
a6ced9d79fe5f8a7e98dd60dc151e4b3dce1b742
15850 F20101123_AABFOI guo_h_Page_041.QC.jpg
a2ee160c7b31a616c9496fb608ef9f18
d3e200c7ea31dd2b50773142f9ca292f85263c49
5571 F20101123_AABFNT guo_h_Page_047thm.jpg
1b2a65119c64db89cf539cc5f75c8056
faa002f6c31053758ac1ba6b73ce99d2616ae3a9
1784 F20101123_AABFOJ guo_h_Page_035.txt
4e67990b6093ae9e656fca5cf98645f1
4a16d3aab7fb5120e7c031fcea3b545e59fe17de
6212 F20101123_AABFNU guo_h_Page_051.pro
e9b6cb08049dd3b102b04666ee962d2f
c5b9322ad4b868bfc0d663247611d2e5f143eec1
47473 F20101123_AABFOK guo_h_Page_083.pro
33cdea3c993a37904daea4aa44084925
cbacaf2b0d718ba4a8d046be3f59cb840f43d7d0
2006 F20101123_AABFNV guo_h_Page_111.txt
e56b9673b516ffbd1cf8d174e0609044
11f2bb0ff26a3d874314797bc1d44d299cbf1330
6023 F20101123_AABFOL guo_h_Page_087thm.jpg
30474779e754f684a043e86df360e5e7
cefdbbf5eeb93f0b9484cd707b71fff4273c6a4a
F20101123_AABFNW guo_h_Page_030.tif
a151fd77d80e42fd29eee9380926b88e
bed6d3d18c9bbe36bd0a37210089848afccfaa5f
F20101123_AABFPA guo_h_Page_111.tif
c6cf834be5ffe47479f4b9d8e58186b4
2b96167df31375d4802f671c03dfa370d8f9775b
40312 F20101123_AABFOM guo_h_Page_103.pro
ddc281aa15a1a15d348f537af636ab2f
6719e5ea86160f2ce289c5976b54189a0abe417f
F20101123_AABFNX guo_h_Page_117.tif
e9cf91ca76ff63152afab49ad2157040
90aa163d89f1a8b4eed1f2e1b4aaa2b8e9059ae6
2062 F20101123_AABFPB guo_h_Page_090.txt
d92b1e282fb371a1131578a4ce4c17ef
e71b758d3085ac971d60c0ea81157d75db1a93c0
4976 F20101123_AABFON guo_h_Page_056thm.jpg
d45b8d568fd63da2ad789a9fff42639f
37880be2e19fa254c56d5724357cb899cd206f0b
26546 F20101123_AABFNY guo_h_Page_033.pro
1d11d9675aa228b3bee114a4b343fe33
7c3b53a418415f736a151b960cce7db406a59ed4
11389 F20101123_AABFPC guo_h_Page_078.pro
cd15b2ad5293790f2616929c3b05b04b
4fb148f19a04667ec42f2d7910db08c995e19cb9
1850 F20101123_AABFOO guo_h_Page_066.txt
1084c16ce71d257c59a16d35dcdc1017
5993366d014194ee8e4a4fcbfed5f17e81b65fd7
6223 F20101123_AABFNZ guo_h_Page_072thm.jpg
4e5f1f8e29169c823ee26df89d20082b
fa0bfded13919da11f56cad166c5d25c265125d1
1424 F20101123_AABFPD guo_h_Page_056.txt
c9cfca69fd05eb2f307358bca4e3d309
9e4e60379494c2a3e127ee22f1409bca21e2d8c2
21672 F20101123_AABFPE guo_h_Page_066.QC.jpg
887a07f15ee1632c584d06b93dc01d57
558aeaf0bb2f0c483f4a3177603a7f37e8707ea0
24050 F20101123_AABFOP guo_h_Page_026.jpg
2eed1d5e4fec28f2967132c602d5500c
b7a63104d5a6de23bb5e8c0801860936c1d72cd6
F20101123_AABFPF guo_h_Page_067.tif
a437be9039184ab3fb27e4e1d26d7089
c11570adaa42b307a93025b26df23647ae8cfcb5
F20101123_AABFOQ guo_h_Page_028.tif
8bd81b856427877eaf5df096385c2038
07c95ff431bd4de1a94dcfe63fee9caca1f23ff5
F20101123_AABFPG guo_h_Page_022thm.jpg
fa109c2c0e9f417ce4775ecf5430ab81
70515b6655445285ea6dc750cc53225160eb5e0a
26159 F20101123_AABFOR guo_h_Page_115.QC.jpg
494550bbf2c1b69742fc5735c3d3383b
72cc45f5ba4daac23770b9698750f7495664481f
F20101123_AABFPH guo_h_Page_063.tif
85583a8c4ab06de3f4eb7f63939149e2
15ff72114d82710c8565acd87ba1da26eedb1db6
8860 F20101123_AABFOS guo_h_Page_075.QC.jpg
08f447ce51814d69fb00b284867428d2
78d023e975f56d068652e2f50c761ac9f6307a64
8190 F20101123_AABFPI guo_h_Page_104.QC.jpg
efb569180b6aae714694abd964b1e54e
381465197c40effbd2b1f3fd998ac5a8081ef2c9
F20101123_AABFOT guo_h_Page_110.tif
af8bc91090e2b67518dd91d79b7c61fd
2cf6438732efc3737f96c4e47500ddd49d784950
8044 F20101123_AABFPJ guo_h_Page_107.QC.jpg
54b03dfc43dab6bcec964ee2b35b08b0
adbddc42940bf0a186a20d584184c88e2cb808c3
898522 F20101123_AABFOU guo_h_Page_042.jp2
060676133384a4076131f35639e6153b
211ff78c72c659969d00080bacddd52444d0655a
55501 F20101123_AABFPK guo_h_Page_023.pro
a75e30dec4ca38f332d2010b6e0fe348
5b64a8ca848e586d23cb70b09b12520f3bad883b
3381 F20101123_AABFOV guo_h_Page_024thm.jpg
9d5a4e21c2f7a4ceffb3d5b9fa3715eb
8f9140060e3c55ef60ce50d1275ab7775bfd1c5e
F20101123_AABFPL guo_h_Page_009.txt
6921a5bc9a4f2964eebe7dae17d448a6
7541f66bb7cd98b09030c7d432bce0baa6169309
6019 F20101123_AABFOW guo_h_Page_012thm.jpg
bac51597d2d4d98a9dfa64abc87e4c8e
0b9cb2c2efbaaa0b96d1abdb3e606d4183131981
946831 F20101123_AABFPM guo_h_Page_054.jp2
823c1ef644e63b888d0467a1ec1a7742
7424a0e676cde60fcaf7721692438aece69ddd63
31016 F20101123_AABFOX guo_h_Page_067.pro
8f4f14d1961e62b67fdb7b8648f0048d
a3844625803f9d4e800f70f5c56de95aca1724e6
791616 F20101123_AABFQA guo_h_Page_085.jp2
cb53296280c6b30d729218361f4f7a46
da650b2cfcd30cd212179a78dba950666081fc8a
954614 F20101123_AABFPN guo_h_Page_011.jp2
e80242c94eccb32830268970fb690ded
d8fe1fb1c3d67249ffbf61f80a94e1e20fb608ff
2158 F20101123_AABFOY guo_h_Page_022.txt
979d5996794b3d6f7723201468b83f8e
5ea4f5be48d6692ba0f32925a0b8982012f5c298
551524 F20101123_AABFQB guo_h_Page_058.jp2
2ad5260532c3e8e7782c7d7b437bed91
e57f2fdcb5e671126d55fb97a32d492362220a70
658422 F20101123_AABFPO guo_h_Page_099.jp2
1f343bcd0d46842d9e576d55a1f5696f
d60bb82abd678fdc60c76009dc74a1c630fc80d9
80441 F20101123_AABFOZ guo_h_Page_057.jpg
a51ecc40a58575b49de9b4bb96362f52
3f5573ec3b68f28d00aa7f181614aefe6bf590ed
49879 F20101123_AABFQC guo_h_Page_012.pro
75c38c6d06132cebf489d25e8077ab55
760d2d9e9690c934529e218ee7d96c522a485b6e
15432 F20101123_AABFPP guo_h_Page_055.QC.jpg
56c98e83ff2e76c603599e06d37a8403
535d981394a3af8b028bd013dd459a5ba5a4db79
F20101123_AABFQD guo_h_Page_047.txt
0c873a655e2ec2833fce076d6016f9d2
fda8fb5db3dbfe5d92128e8fd5f440f499a7bfa0
1699 F20101123_AABFQE guo_h_Page_046.txt
329eff13fbb4bf734367d11ffe2e3a84
87af0b90e40d01e6e77ac7d653e81ffe51d48db6
5963 F20101123_AABFQF guo_h_Page_028thm.jpg
ae0d2848e00718686948680847247535
8ba32b07f522464d28088f96e117161ad00d6c8e
870157 F20101123_AABFPQ guo_h_Page_047.jp2
290aef7ddafd2c415004c6395f7119f7
7bc8996abd7a1000ac3838b7caa62daa57b31544
42692 F20101123_AABFQG guo_h_Page_011.pro
45bba76d41335c4ddedba8b79c89f0a7
c59d260802ba77c5b628da544660b7cc395f4e50
38581 F20101123_AABFPR guo_h_Page_035.pro
ac34cb1eff30ead01c6c29ffcd4189c4
d647fb21bf616647beb61ff5e8944fce6a3377f6
54767 F20101123_AABFQH guo_h_Page_040.pro
e85567d34c0027328a601015209ba988
1fd562a5c3741c93591a4100d7cbf312bcadf906
F20101123_AABFPS guo_h_Page_098.tif
22978fbd79e3f5a001dd5ffcc4fdc07d
896161a9299640b10a55b778c1a33a092c9babc9
5477 F20101123_AABFQI guo_h_Page_032thm.jpg
e191bf8c5af206c626e445dca46d8e32
bf66670140ecab33a08586bb0cac85035ad71ad9
70149 F20101123_AABFPT guo_h_Page_061.jpg
60bacc72501cf3f66fa333e2377adb09
f83ec68eb93ba8df35e134e246ffb3a076dabb09
F20101123_AABFQJ guo_h_Page_002.tif
f25c849724bd6827cf33a9e3acd35fe4
4cb6402a920a6509fa6a8fbc35ea7828b267d08f
434839 F20101123_AABFPU guo_h_Page_020.jp2
eb7219728de450644a5d8d1c80bd6ea2
22dc6ac74bdb95fb08f504bb0d47e86edb08c55a
68535 F20101123_AABFQK guo_h_Page_082.pro
ab2e2c2cca7a5482211048206a67ff43
edc1a5c67fa34d11dd81218e99416aa0a30077ef
F20101123_AABFPV guo_h_Page_094.jp2
fa352c4e69fa13a26f5ca44ab067c0e7
6414a4a3262cd3dc2e72f9e877cf11b030df364a
F20101123_AABFQL guo_h_Page_029thm.jpg
a6d2fa150d528374314f7315bbc6fa8f
70d8a0c6073d64134ec025b4328f524fd453e17c
25797 F20101123_AABFPW guo_h_Page_100.pro
2b7838116044aeef62ebc1cc38c8f2bc
c21209b93d0407239f1d6c7a50bbb88731e24f9e
1357 F20101123_AABFRA guo_h_Page_002thm.jpg
83a31c0a16768c2e7381555d7c03fffd
c6323726dc5cdaffacce9db389c0e4f589b0c088
22396 F20101123_AABFQM guo_h_Page_036.QC.jpg
a9f0e1a61b7261417d09f859b329b665
f8d7be7bd9e934780b4f3c40dcb4d366994059f9
2045 F20101123_AABFPX guo_h_Page_072.txt
65bbab44e433e6c9ea1bafebc467b326
47de3f41889308c2d198cb2a0a1b553bd76d4226
4063 F20101123_AABFRB guo_h_Page_098.pro
a330081b90caa189ffd43a735cd09935
984d8c10dcb27a7d7ec05160507ccd8dd50803c0
F20101123_AABFQN guo_h_Page_034.tif
b081e34672df888edc834a6c631f6641
e2aad0212449cbcbdc8f28d6125c7cb5b14cf6d6
5906 F20101123_AABFPY guo_h_Page_020.pro
2c707ea8801431d9d68b81374e6ed54d
28c338064b8afa1eb3ff29dbe9b0045eaaa4a951
267141 F20101123_AABFRC guo_h_Page_098.jp2
6755200e2e010544bd7dfd57f9045828
072aa44891ad2f36bd69e73d6a3d7b16c50b7c94
F20101123_AABFQO guo_h_Page_085.tif
4f5ca582c2669bdacb15011df1a7f45b
863d4ecb252a2336fa8ac241b231fc31288ccb21
F20101123_AABFPZ guo_h_Page_083.txt
de4ccad61cd75cade97fcd8897e399c2
0c5fa9d71bbea3f2a1ffcc5511b99165cac6ca1a
2580 F20101123_AABFRD guo_h_Page_075thm.jpg
e452290ce9b333f7427e610aa9bfdc97
f4c2eaa20bb9da535be47ea817c36b5e9d69d26a
613063 F20101123_AABFQP guo_h_Page_055.jp2
ff104d5378da0c2f305a5300a0c4888b
6fed6436f203f7ba40f4d35426f3cc4e16758097
829233 F20101123_AABFRE guo_h_Page_035.jp2
48fce42be63e184dfb66718ae1d2926b
3a6b2f661709a689c69fc8e1ebf1e5ae5244f766
27376 F20101123_AABFQQ guo_h_Page_078.jp2
916f70447771f82f2cd7ed30ab84a9ff
465605d0576324f4b097fec15e386790f5df90ad
445 F20101123_AABFRF guo_h_Page_116.txt
9463c676e6e4b7409a94eafe5ed8abfc
9ac0245fd9a380ce4dc2f9481ff7cbade948ce68
23898 F20101123_AABFRG guo_h_Page_058.pro
2ae1cf4f9296b3edff68afc005b4d160
cddb8a193a7ca8fa077edaa4f013d63f6410293b
49739 F20101123_AABFQR guo_h_Page_043.pro
825faf8fd9c0423427956216038dca4f
efe59ffcac8ccf2bd082def025f7fda091bedd9e
2257 F20101123_AABFRH guo_h_Page_117thm.jpg
cffb737814d03de8139f55238dabbed7
68ce4aff7efeea6f9185a18869df5946b714070a
43459 F20101123_AABFQS guo_h_Page_068.jpg
54ceeb03a75f6ebd9ac35c571b204825
25f4516b98df7d7fa41e0b4785bef462dec0ae79
6661 F20101123_AABFRI guo_h_Page_115thm.jpg
3a67edea170bf3f418d857f06b25628f
e8dd58127ab38cf2c26b89b3ede77a19d1bbea59
22272 F20101123_AABFQT guo_h_Page_003.jp2
8af9904a0461c78ce0a69051d31c4d75
9027e090e4123945434537621005e69da9e47a7b
5191 F20101123_AABFRJ guo_h_Page_035thm.jpg
658efa1916533256f67e2d5e423008e8
6a0ffbd167ce828203f5ea2e9973d5da5fcfd8ee
10111 F20101123_AABFQU guo_h_Page_050.QC.jpg
9142faec64577a78267c7732e1abbcb0
d785d9466e9622090ad2dfde7fab1cb83b26a2b8
F20101123_AABFRK guo_h_Page_088.txt
bfc25a16e94a6342d9ceb7ee9592b95a
44f786d05bd10b523d4a7cb604deb7fa139d17f3
F20101123_AABFQV guo_h_Page_096.tif
7036c2d54f54d8c0836aa2f5d738dd18
391bd85b7daae2e7f305f747a0e1398d35fd938a
854866 F20101123_AABFRL guo_h_Page_070.jp2
30a19b70d45e0cc68f9089ed584e2001
4c0208f5b887d91a052d05dbab8ce4d6c84f4e97
51364 F20101123_AABFQW guo_h_Page_090.pro
731a25ba5495666c84d90eac4357e06c
480adcc6ba4c599cdd7ff8066bdda4abda61ac86
8513 F20101123_AABFRM guo_h_Page_106.QC.jpg
a1736170a03d08be61f28c7b2c725c85
9b46008582c14e50917de11c13f1f7c4c33be549
6348 F20101123_AABFQX guo_h_Page_116.QC.jpg
3bcc7d2e55a641b2fe86fb110d5a043e
021178f300ecf6f00ba4ecd620525dbe493af89e
18049 F20101123_AABFSA guo_h_Page_064.QC.jpg
fd28bf683c61004396773af0fd115460
d292fa2a7378211720fb1bdd846a0056e2c164d6
F20101123_AABFRN guo_h_Page_045.jp2
6ec2efded7a213212bc13d4dced5c86e
4a5739a10d4e5e520ada0af19689b79335c89737
5945 F20101123_AABFQY guo_h_Page_092.pro
a06d0730ae3043910e25fd20c7e5fb77
bc80cf9179460afcf5143349a7bd061ccd548c3a
48490 F20101123_AABFSB guo_h_Page_073.pro
aabb03de8cb369c8b627976725792968
bac8081eaaf48802c53694b1ad11f82a75ab6d43
25784 F20101123_AABFRO guo_h_Page_082.QC.jpg
d612248b410e5b0589afcc62dffa37df
70fb9a0cc3147c4b7733559125844d196c97bdac
2072 F20101123_AABFQZ guo_h_Page_073.txt
0b23e7d011f23dab84fcce3e9b6fc2db
ba30b1d1bd7c6f09b8b2fde1b6be370f20fa5c48
1041 F20101123_AABFSC guo_h_Page_002.pro
683fbf43a31c0bb116ff492d697923f7
c5006112a126af0fb7cc653d8409f70bfecc6339
917228 F20101123_AABFRP guo_h_Page_096.jp2
23a0f9798d83d8ed92007cf7a567dbb3
73b4da349e6e11314f613e6959ed9841282c73f0
15786 F20101123_AABFSD guo_h_Page_106.pro
074f0f39f8305f415b507ce1429d5f3d
acb4c08f78ce0d20ac51e5b3910eb4e992402453
53273 F20101123_AABFRQ guo_h_Page_016.jpg
60b7804902f83573c0b7dfd5c36196bf
8f4f95a19df56024b1b5ab301128772daf33edc6
2077 F20101123_AABFSE guo_h_Page_025.txt
4aba9f20a0bc1c8d57754733eaa0d775
e9e25c0542a4d3d7604da9df2c5929d00cc344a3
F20101123_AABFRR guo_h_Page_039.tif
da06307a8b10ea30f8fc1c437879099e
83256be599d766ece3c37f30addd2935a84d87b4
F20101123_AABFSF guo_h_Page_082.tif
682cc6357b374034a12421e57541605f
d167fb14153564e1398472f2c042453878d18f0a
44524 F20101123_AABFSG guo_h_Page_053.pro
7af48e208f29d9ed207cc87417ffee77
0a5e3d51ac5bf1c8b2d155c7881b9cb4050fd2ac
9582 F20101123_AABFRS guo_h_Page_003.jpg
d7890dc644a1c1969e237d094d3fadae
99770c0a48cf0d3f3fa352ae275494eba116fa7b
66750 F20101123_AABFSH guo_h_Page_081.jpg
2811d62e9414065baa162d1b06d6ee2f
f08ab4bde77e0ab65c98de8ace2143c479f71771
1552 F20101123_AABFRT guo_h_Page_085.txt
a656c6d3b51e5043f710ca5f0570d490
0e612159bfa5d4957c4bbb841966882d9b69c37c
9048 F20101123_AABFSI guo_h_Page_116.pro
3cb65fedae160a6706a3e8ecbfa809ee
28aec0c6013798e055315db7996c96ee0534889c
57236 F20101123_AABFRU guo_h_Page_065.jpg
27f4e4029607db32b3e0684b1d465d42
b318d5ebeee3a80db88b1db44e0681a1eb055acc
3711 F20101123_AABFSJ guo_h_Page_062thm.jpg
8ad67dbfb12b0de4dfb00491dee5e0e2
0cc3e95c9c0b04395536026bf5b089eade9ac356
F20101123_AABFRV guo_h_Page_016.tif
4c827c360d0104f2d889387734d90e2e
d5fb727bacab5faf28f8c334ee427590f5182547
F20101123_AABFSK guo_h_Page_073.tif
26b90e87f869f3b7829b88da17a332ac
92c1aa6a93a8abe8bb27a9557901743164476437
F20101123_AABFRW guo_h_Page_072.tif
b4230ffacbd1b7b11a66ad51e1273ed0
5bef30cd8e250f0a37e0e85f5e6e4238705197e3
26465 F20101123_AABFSL guo_h_Page_112.QC.jpg
204b7987f2fc0cf40918479f46b2aeef
2e35d55cf2b60e99304f7271b44eeac9b96bedab
5438 F20101123_AABFRX guo_h_Page_011thm.jpg
f1c7393c7de7b1e3ef1d57b5cc62ae46
4d8e6e5148243f55330d59e45bedfa4739697814
44112 F20101123_AABFTA guo_h_Page_005.jpg
0444f165a7bfdb2d7818b26ac9347d7b
89c4adb053ef9a63ac2cfaa84b3714b2846b6e14
60042 F20101123_AABFSM guo_h_Page_084.jpg
ed9eeb037f6af7582621158d1bf42e0c
552136cf0ed7ef9b2117f51b6fe69646c2ac265f
22422 F20101123_AABFRY guo_h_Page_059.QC.jpg
91ef37a319a78dedafd4457f8e6dbdbd
5dc697a7a297b678fe759f829183672db7b6734a
F20101123_AABFTB guo_h_Page_075.tif
cb70299c3bc93f0e61c84e7a2f793013
4386a4055d7c7766ff299ecbb5823020d96ce894
F20101123_AABFSN guo_h_Page_090.jp2
1ebd948402c26479b49b88a8d347bc12
27c74ddf95034bbf9b2711696dafaafdf5d6043c
F20101123_AABFRZ guo_h_Page_061.tif
c97718d493e2900626700c6d90ad9f2e
691da8419164c1706881f21dbc5fb2268707d8c9
3299 F20101123_AABFTC guo_h_Page_007thm.jpg
43da6fdfff98dfc9bb364456e25be879
72cdb472f557b45c54b83898a321ecb6868b61af
88735 F20101123_AABFSO guo_h_Page_114.jpg
0a1b8032322552805ba2526d8ebf90d9
31aa3642869df8f4dcb2e6307354c8e279c6ab59
217 F20101123_AABFTD guo_h_Page_107.txt
83caac573cb7e3194dfc2aad0eeea10a
be3c3e182180727ccae04774f8a7f354492fc6e4
23106 F20101123_AABFSP guo_h_Page_083.QC.jpg
c3172ef6d1e46d941b622f0258daa8c9
588963f46d5d72c609a4c2b3647ced5d622d48eb



PAGE 1

DIFFEOMORPHIC POINT MA TCHING WITH APPLICA TIONS IN MEDICAL IMA GE ANAL YSIS By HONGYU GUO A DISSER T A TION PRESENTED TO THE GRADUA TE SCHOOL OF THE UNIVERSITY OF FLORID A IN P AR TIAL FULFILLMENT OF THE REQUIREMENTS F OR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORID A 2005

PAGE 2

Cop yrigh t 2005 b y Hongyu Guo

PAGE 3

T o m y daugh ter Alicia.

PAGE 4

A CKNO WLEDGMENTS I will b e forev er indebted to m y advisor, Dr. Anand Rangara jan, for his in v aluable guidance and supp ort. I am v ery luc ky to ha v e the opp ortunit y to w ork with him. I am v ery grateful for his appreciation and condence in m y mathematical talen t. Although he is an outstanding researc her in the computer vision and medical imaging eld with man y distingushed con tributions, he is v ery h um ble and do wn to earth. During the y ears of m y study I could kno c k at his do or an y time I needed him and he w ould put do wn his w ork and discuss sparkling ideas with me. W orking with him has b een truly a w onderful and rew arding exp erience. I w ould lik e to express deep est thanks to Dr. Aruna v a Banerjee, Dr. Jorg P eters and Dr. Baba V em uri of the Departmen t of CISE, Dr. Y unmei Chen of the Departmen t of Mathematics and Dr. Haldun Aytug of the Departmen t of Decision and Information Sciences for serving on m y sup ervisory committee. I thank them for their in v aluable criticism and advice. I wish to extend m y gratitude to Dr. Sarang Joshi at the Univ ersit y of North Carolina at Chap el Hill and Dr. Lauren t Y ounes at Johns Hopkins Univ ersit y They are the coauthors of a b o ok c hapter and sev eral pap ers that I ha v e written and this w ork in m y dissertation is greatly b enetted from their previous w ork on dieomorphic landmark matc hing. I w ould lik e to thank Dr. Eric Grimson, the director of MIT AI Lab, for the helpful discussion and v aluable commen ts when he visited the Univ ersit y of Florida. I wish to extend sp ecial thanks to Dr. Alp er for an enligh tening discussion on triangulation and surface reconstruction. iv

PAGE 5

My thanks also go to ev ery one of m y fello w studen ts at the Cen ter for Computer Vision, Graphic sand Medical Imaging at the Univ ersit y of Florida, particularly Bing Jian, San thosh Ko dipak a, Jie Zhang, Adrian P eter, Zhizhou W ang, Eric Sp ellman, F ei W ang, Tim McGra w and Xiaobin W u. I get constan t help from them ev ery da y I am indebted to Dr. John M. Sulliv an at the Univ ersit y of Illinois at UrbanaChampaign for gran ting p ermission to use his graphics image of the Klein b ottle. I w ould lik e to thank m y wife Y anping for her lo v e and supp ort and I thank m y daugh ter Alicia for b eing a w onderful daugh ter. Alicia and I enjo y ed w orking together on man y w eek ends. While she w as fascinated b y the rotating shap es of computer graphics I displa y ed with MA TLAB, I admired her b eautiful and creativ e colorings and dra wings. This researc h is supp orted in part b y the National Science F oundation gran t I IS 0307712. v

PAGE 6

T ABLE OF CONTENTS page A CKNO WLEDGMENTS . . . . . . . . . . . . . . iv LIST OF T ABLES . . . . . . . . . . . . . . . . viii LIST OF FIGURES . . . . . . . . . . . . . . . . ix ABSTRA CT . . . . . . . . . . . . . . . . . . xi CHAPTER1 INTR ODUCTION . . . . . . . . . . . . . . . 1 2 PREVIOUS W ORK . . . . . . . . . . . . . . 17 2.1 Thin Plate Splines (TPS) . . . . . . . . . . . 18 2.2 Bo okstein's Application to 2D Landmark W arping . . . . 19 2.3 More on Splines . . . . . . . . . . . . . . 21 2.4 Repro ducing Kernel Hilb ert Space (RKHS) F orm ulation . . . 23 2.5 The F olding Problem of Splines . . . . . . . . . 24 2.6 Imp osing Restriction on the Jacobian . . . . . . . . 25 2.7 The Flo w Approac h . . . . . . . . . . . . . 25 2.8 Corresp ondence and Softassign . . . . . . . . . . 28 2.9 Distance T ransforms . . . . . . . . . . . . 28 2.10 Implicit Corresp ondence . . . . . . . . . . . 29 2.11 Shap e Con text . . . . . . . . . . . . . . 31 2.12 Activ e Shap e Mo dels . . . . . . . . . . . . 31 2.13 Deterministic Annealing Applied to EM Clustering . . . . 32 2.14 Statistical Shap e Analysis on Dieren tiable Manifolds . . . 34 2.15 Distance Measures from Information Theory . . . . . . 36 3 DIFFEOMORPHIC POINT MA TCHING . . . . . . . . 42 3.1 Existence of a Dieomorphic Mapping in Landmark Matc hing . 42 3.2 Symmetric Matc hing due to Time Rev ersibilit y . . . . . 44 3.3 A Theoretical F ramew ork for Dieomorphic P oin t Matc hing . 46 3.4 A Dieomorphic P oin t Matc hing Algorithm . . . . . . 49 3.5 Applications to 2D Corpus Callosum Shap es . . . . . . 53 3.6 Applications to 3D Shap es . . . . . . . . . . . 55 3.6.1 Exp erimen ts on Syn thetic Data . . . . . . . 59 3.6.2 Exp erimen ts on Real Data . . . . . . . . . 61 vi

PAGE 7

4 TOPOLOGICAL CLUSTERING AND MA TCHING . . . . . 72 4.1 F undamen tals of T op ological Spaces . . . . . . . . 72 4.2 Kohonen Self-Organizing F eature Map (SOFM) . . . . . 76 4.3 T op ological Clustering and Matc hing . . . . . . . . 79 4.3.1 Wh y: the Need for T op ology . . . . . . . . 79 4.3.2 Ho w: Graph T op ology . . . . . . . . . . 80 4.4 Ob jectiv e F unction and the Algorithm . . . . . . . 84 4.5 Prescrib ed T op ology . . . . . . . . . . . . 85 4.5.1 Chain T op ology . . . . . . . . . . . . 85 4.5.2 Ring T op ology . . . . . . . . . . . . 85 4.5.3 S 2 T op ology . . . . . . . . . . . . . 85 4.6 Arbitrary T op ology . . . . . . . . . . . . . 89 5 CONCLUSIONS . . . . . . . . . . . . . . . 98 5.1 Con tributions . . . . . . . . . . . . . . 98 5.2 F uture W ork . . . . . . . . . . . . . . . 99 REFERENCES . . . . . . . . . . . . . . . . . 100 BIOGRAPHICAL SKETCH . . . . . . . . . . . . . . 106 vii

PAGE 8

LIST OF T ABLES T able page 3{1 Mo died Hausdor distance of the matc hing p oin t sets. . . . . 57 3{2 Geo desic distances and mo died Hausfor distances . . . . . 58 3{3 Matc hing errors on syn thetic data with dieren t noise lev els . . . 60 3{4 Limiting v alue of determined b y the n um b er of clusters . . . 61 3{5 Jensen-Shannon div ergence for v arious pairs of shap es . . . . 70 3{6 Hausdor and mo died Hausdor distance for v arious pairs of shap es 71 viii

PAGE 9

LIST OF FIGURES Figure page 1{1 Grids in R 2 (a) b efore transformation and (b) after transformation . 9 1{2 Klein b ottle immersed in R 3 . . . . . . . . . . . . 9 1{3 (a) The template image and (b) The reference image . . . . 10 1{4 Space deformations . . . . . . . . . . . . . . 10 1{5 Am biguit y of landmark corresp ondence . . . . . . . . 13 1{6 Tw o shap es of a ribb on and the asso ciated landmarks . . . . 15 1{7 Misin terpretation of the shap e . . . . . . . . . . . 16 2{1 Deformation of the thin plate . . . . . . . . . . . 38 2{2 Landmark displacemen ts . . . . . . . . . . . . 38 2{3 Thin-plate Spline in terp olation . . . . . . . . . . . 39 2{4 Dieomorphic in terp olation . . . . . . . . . . . 39 2{5 Tw o distance transformed images of three landmarks . . . . 40 2{6 Lev el sets of t w o distance transformed images . . . . . . 41 3{1 Existence of a dieomorphic mapping . . . . . . . . . 44 3{2 Asymmetry of the matc hing . . . . . . . . . . . 45 3{3 P oin t sets of nine corpus callosum images. . . . . . . . 56 3{4 Clustering of the t w o p oin t sets. . . . . . . . . . . 56 3{5 Dieomorphic mapping of the space. . . . . . . . . . 56 3{6 Matc hing b et w een the t w o p oin t sets. . . . . . . . . . 57 3{7 Ov erla y of the after-images of eigh t p oin t sets with the nin th set. . 57 3{8 Tw o p oin t sets of hipp o campal shap es . . . . . . . . 59 3{9 Matc hing errors on syn thetic data for dieren t n um b er of clusters . 65 3{10 Tw o hipp o campal shap es . . . . . . . . . . . . 66 ix

PAGE 10

3{11 Deterministic annealing in the clustering pro cess . . . . . . 67 3{12 Limiting v alue of determined b y the n um b er of clusters . . . 68 3{13 Clustering of the t w o hipp o campal shap es . . . . . . . 69 4{1 Dieren t top ological spaces . . . . . . . . . . . . 77 4{2 Image con tours of t w o hands . . . . . . . . . . . 80 4{3 Clustering of t w o hands . . . . . . . . . . . . . 81 4{4 Corresp ondence . . . . . . . . . . . . . . . 81 4{5 Finite top ology . . . . . . . . . . . . . . . 82 4{6 T op ological clustering and matc hing of t w o hands . . . . . 86 4{7 Corresp ondence with top ology constrain t . . . . . . . . 86 4{8 T op ological clustering of corpus callosum shap es . . . . . . 87 4{9 Corresp ondence in top ological clustering of corpus callosum shap es . 88 4{10 Sphere top ology . . . . . . . . . . . . . . . 89 4{11 T op ological clustering of hipp o campus with S 2 top ology: the rst set 90 4{12 T op ological clustering of hipp o campus with S 2 top ology: the second set 91 4{13 Graph top ology of hipp o campus shap e through learning . . . . 93 4{14 T op ological clustering of hipp o campus shap e: the rst p oin t set . 94 4{15 T op ological clustering of hipp o campus shap e: the second p oin t set . 95 4{16 Fish shap es . . . . . . . . . . . . . . . . 95 4{17 Graph top ology for the sh shap e with 4 nearest neigh b ors . . . 96 4{18 T op ological clustering and matc hing of the sh shap es . . . . 97 x

PAGE 11

Abstract of Dissertation Presen ted to the Graduate Sc ho ol of the Univ ersit y of Florida in P artial F ulllmen t of the Requiremen ts for the Degree of Do ctor of Philosoph y DIFFEOMORPHIC POINT MA TCHING WITH APPLICA TIONS IN MEDICAL IMA GE ANAL YSIS By Hongyu Guo August 2005 Chair: Dr. Anand Rangara jan Ma jor Departmen t: Computer and Information Science and Engineering Dieomorphic matc hing of unlab eled p oin t sets is v ery imp ortan t to non-rigid registration and man y other applications but it has nev er b een done b efore. It is a v ery c hallenging problem b ecause w e ha v e to solv e for the unkno wn corresp ondence b et w een the t w o p oin t sets. In this w ork w e prop ose a join t clustering metho d to solv e for a sim ultaneous estimation of the corresp ondence and the dieomorphism in space. The cluster cen ters in eac h p oin t set are alw a ys in corresp ondence b y virtue of ha ving the same index. During clustering, the cluster cen ter coun terparts in eac h p oin t set are link ed b y a dieomorphism and hence are forced to mo v e in lo c k-step with one another. W e devise an ob jectiv e function and design an algorithm to nd the minimizer of the ob jectiv e function. W e apply the algorithm to 2D and 3D shap es in medical imaging. W e further prop ose to use a graph represen tation for the shap e top ology information. Results are giv en for prescrib ed top ologies lik e c hain top ology ring top ology { whic h are v ery common in dealing with 2D con tour shap es { and gen us zero closed surface top ology in 3D. W e also in v estigate the top ology problem in general and the learning of top ology with a nearest neigh b or graph. xi

PAGE 12

CHAPTER 1 INTR ODUCTION This w ork is a rst time eort to study the dieomorphic p oin t matc hing problem with unkno wn corresp ondence. In this c hapter w e in tro duce the need for an eectiv e dieomorphic p oin t matc hing algorithm and the bac kground kno wledge of the shap e matc hing eld. W e giv e some fundamen tal denitions and theorems from dieren tial top ology whic h are essen tial in the dev elopmen t of the follo wing c hapters. W e analyze the nature and dicult y of the corresp ondence problem. Shap e analysis using to ols of mo dern dieren tial geometry and statistical theory has far and wide applications in computer vision, image pro cessing, biology morphometrics, computational anatom y biomedical imaging, and image guided surgery as w ell as arc heology and astronom y Shap es pla y a fundamen tal role in computer vision and image analysis and understanding. In general terms, the shap e of an ob ject, a data set, or an image can b e dened as the total of all information that is in v arian t under certain spatial transformations [ 66 ]. Shap e matc hing and corresp ondence problems arise in v arious application areas suc h as computer vision, pattern recognition, mac hine learning and esp ecially in computational anatom y and biomedical imaging. Shap e matc hing b ecomes an indisp ensable part of man y biomedical applications lik e medical diagnosis, radiological treatmen t, treatmen t ev aluation, surgical planning, image guided surgery and pathology researc h. Shap es ma y ha v e man y dieren t represen tations. They can b e represen ted with the in tensities of pixels of an image, whic h is a function dened in a region of 2D or 3D space, or they can b e represen ted with p oin t sets, curv es or surfaces. In this dissertation, w e fo cus on the p oin t represen tation of shap es. P oin t represen tation of image data is widely used in all areas and there is a h uge amoun t of 1

PAGE 13

2 p oin t image data acquired in v arious mo dalities, including optical, MRI, computed tomograph y and diusion tensor images [ 18 24 ]. The adv an tage of p oin t set represen tation of shap es, as opp osed to curv e and surface represen tations is m ultifarious and that is wh y w e fo cus on p oint shap e matching instead of curv e or surface matc hing [ 64 68 67 ]. The p oin t set represen tation is compact in storage. The computational time when using p oin t set represen tation is dramatically reduced as opp osed to when image in tensities are used. The p oin t set represen tation of shap es is univ ersal and homogeneous. It do es not require the prior kno wledge ab out the top ology of shap es. It has the capabilit y to fuse dieren t t yp es of features in to a global, uniform and homogeneous represen tation. A p oin t set represen tation of shap es is esp ecially useful when feature grouping (in to curv es and the lik e) cannot b e assumed. Statistical analysis on p oin t set shap es is straigh tforw ard, as demonstrated in Co otes et al. [ 18 ] using activ e shap e mo dels. The recen t w ork of Glaunes et al. [ 27 ] is a b old step forw ard to generalize the p oin t set shap es to general Radon measures and distributions in the sense of Sc h w artz generalized functions, in order to mo del shap es represen ted b y a mixture of p oin ts and submanifolds of dieren t dimensions (curv es and surfaces). Although this is viable in theory it b ecomes unpractical when it comes to applications. In their exp erimen ts dealing with a mixture of p oin ts and a curv e, they used the tec hnique of resampling the curv e. That comes bac k to the p oin t set represen tation of the curv e itself. Moreo v er, b y doing so, new free parameters, i.e., the relativ e w eigh t of measure b et w een p oin ts and the curv es, is in tro duced, whic h can b e arbitrary P oin t shap e matc hing is ubiquitous in medical imaging and in particular, there is a real need for a turnk ey non-rigid p oin t shap e matc hing algorithm [ 18 24 19 ]. P oin t shap e matc hing in general is a dicult problem b ecause, as with man y other problems in computer vision, lik e image registration and segmen tation, it

PAGE 14

3 is often ill-p osed. W e try to mak e abstractions out of the practical problems to form ulate precise mathematical mo dels of the problem. P oin t matc hing can b e view ed in the con text, or out of the con text of image registration. In the con text of image registration, the p oin ts are view ed as feature p oin ts in the image. The corresp ondence can b e kno wn or unkno wn. When the cardinalit y of the t w o p oin t sets is the same and when the corresp ondence is kno wn, w e call this the landmark matching pr oblem The n um b er of feature p oin ts ma y also b e unequal. In that case w e are dealing with outliers. The dicult y increases dramatically when the corresp ondence is not kno wn and/or when there are outliers. The p oin t matc hing problem can also exist out of the con text of image registration. In this case, the p oin ts are samples from a shap e and w e ha v e a p oin t represen tation of the shap e. When w e ha v e t w o suc h shap es represen ted b y p oin ts, usually the n um b er of p oin ts in the t w o shap es is dieren t and there is no p oin t-wise corresp ondence. W e w an t to nd the corresp ondence b et w een the t w o shap es. W e call this the p oint shap e matching pr oblem. In the follo wing, w e will rst address the landmark matc hing problem and then the p oin t shap e matc hing problem. W e assume the image domain is d -dimensional Euclidean space R d Usually d = 2 or d = 3. Supp ose w e ha v e t w o images, I 1 : n 1 R and I 2 : n 2 R where n 1 R d and n 2 R d The image registration problems can b e classied in to t w o categories: in tensit y based registration and feature based registration [ 52 ]. In the in tensit y based image registration, w e need to nd a map f : n 1 n 2 suc h that 8 x 2 n 1 ; I 1 ( x ) = I 2 ( f ( x )). This is the ideal registration problem. In feature based registration, w e supp ose w e ha v e t w o corresp onding sets of feature p oin ts, or landmarks, f p i 2 n 1 j i = 1 ; 2 ; :::; n g and f q i 2 n 2 j i = 1 ; 2 ; :::; n g W e need to nd a transformation f : n 1 n 2 suc h that 8 i = 1 ; 2 ; :::; n f ( p i ) = q i In man y applications, w e are required to nd the transformation within some restricted groups, lik e rigid transformations, similarit y transformations, ane

PAGE 15

4 transformations, pro jectiv e transformations, p olynomial transformations, B-spline transformations and \non-rigid" transformations. Dieren t transformation groups ha v e dieren t degrees of freedom, the n um b er of parameters needed to describ e a transformation in the group. This also determines the n um b er of landmark pairs that the transformation can exactly in terp olate. Let us lo ok at some examples. In the t w o dimensional space, where d = 2, a rigid transformation, whic h preserv es Euclidean distance, dened b y 1.1 has 3 degrees of freedom ( '; x 0 ; y 0 ) and cannot in terp olate arbitrary landmark pairs. 264 x 0 y 0 375 = 264 cos sin sin cos 375 264 x y 375 + 264 x 0 y 0 375 (1.1) The landmark pairs to b e matc hed m ust b e sub ject to some constrain ts. That is, they ha v e to ha v e the same Euclidean distance. A similarit y transformation dened b y 1.2 has 4 degrees of freedom ( k ; '; x 0 ; y 0 ) and can map an y 2 p oin ts to an y 2 p oin ts. 264 x 0 y 0 375 = k 264 cos sin sin cos 375 264 x y 375 + 264 x 0 y 0 375 (1.2) An ane transformation dened b y 1.3 has 6 degrees of freedom ( a 11 ; a 12 ; a 21 ; a 22 ; x 0 ; y 0 ) and can map an y 3 non-degenerate p oin ts to an y 3 nondegenerate p oin ts. 264 x 0 y 0 375 = 264 a 11 a 12 a 21 a 22 375 264 x y 375 + 264 x 0 y 0 375 (1.3) A pro jectiv e transformation dened b y 1.4 in term of non-homogeneous co ordinates has 8 degrees of freedom and can map an y 4 non-degenerate p oin ts to an y 4 non-degenerate p oin ts.

PAGE 16

5 264 x 0 y 0 375 = 264 a 11 x + a 12 y + a 13 d 11 x + d 12 y +1 b 11 x + b 12 y + b 13 d 11 x + d 12 y +1 375 (1.4) In three dimensional space, where d = 3, w e write the transformation in a more general form, 266664 x 0 y 0 z 0 377775 = 266664 a 11 a 12 a 13 a 21 a 22 a 23 a 31 a 32 a 33 377775 266664 x y z 377775 + 266664 x 0 y 0 z 0 377775 ; (1.5) where A = 266664 a 11 a 12 a 13 a 21 a 22 a 23 a 31 a 32 a 33 377775 is a 3 3 matrix, whic h represen ts a linear transformation. A rigid transformation requires A to b e orthogonal and has 6 degrees of freedom. A similarit y transformation requires A to b e a similar matrix and has 7 degrees of freedom. An ane transformation allo ws A to b e the most general form and has 12 degrees of freedom and can map an y 4 non-degenerate p oin ts to an y 4 non-degenerate p oin ts. A projectiv e transformation in 3D as dened b y 1.6 has 15 degrees of freedom and can map an y 5 non-degenerate p oin ts to an y 5 non-degenerate p oin ts. 266664 x 0 y 0 z 0 377775 = 266664 a 11 x + a 12 y + a 13 z + a 14 d 11 x + d 12 y + d 13 z +1 b 11 x + b 12 y + b 13 z + b 14 d 11 x + d 12 y + d 13 z +1 c 11 x + c 12 y + c 13 z + c 14 d 11 x + d 12 y + d 13 z +1 377775 (1.6) The term \non-rigid" transformation is often used in a narro w er sense. Although similarit y ane and pro jectiv e transformations do not preserv e Euclidean distance, they all ha v e nite degrees of freedom. In the literature, \non-rigid"

PAGE 17

6 transformation usually refers to a transformation of innite degrees of freedom, whic h can p oten tially map an y nite n um b er of p oin ts to the same n um b er of p oin ts. So w e immediately see a big dierence b et w een nite degree of freedom transformations and non-rigid transformations. Giv en a xed n um b er of landmark pairs to b e in terp olated, the former is easily o v er constrained; the latter is alw a ys under constrained. This is one of the reasons that the non-rigid p oin t matc hing problem is m uc h more dicult. T o nd a unique non-rigid transformation, w e need further constrain ts. W e call this regularization. Tw o desirable prop erties of non-rigid transformations are smo othness and top ology preserving. Let n 1 R d and n 2 R d A transformation f : n 1 n 2 is said to b e smo oth if all partial deriv ativ es of f up to certain orders, exist and are con tin uous. If transformation f : n 1 n 2 preserv es the top ology then n 1 and Img ( f ) = f p 2 2 n 2 j9 p 1 2 n 1 ; p 2 = f ( p 1 ) g ha v e the same top ology A transformation that preserv es top ology means w e will not ha v e tears in space or in the image. A transformation that preserv es top ology is called a home omorphism and its formal denition is: Denition 1. L et n 1 and n 2 b e two top olo gic al sp ac es. A map f : n 1 n 2 is a home omorphism if f is a bije ction; f is c ontinuous; the inverse f 1 is c ontinuous. A smo oth transformation f : n 1 n 2 ma y not preserv e the top ology Namely a smo oth map ma y not b e a homeomorphism. It is easy to see this b ecause a smo oth map ev en ma y not b e a bijection. On the other hand, a homeomorphism ma y not b e smo oth b ecause in the denition, w e only require con tin uit y in b oth f and its in v erse but w e do not require dieren tiabilit y It is strongly desirable that

PAGE 18

7 the transformation is b oth smo oth and top ology preserving. What w e w an t is a die omorphism whic h is dened as follo ws. Denition 2. L et M 1 and M 2 b e dier entiable manifolds. A map f : M 1 M 2 is a die omorphism if f is a bije ction; f is dier entiable; the inverse f 1 is dier entiable. Because the concept of dieomorphism is essen tial to this w ork, w e w ould ha v e a little more discussion here in order to clarify some common misconceptions ab out this concept. It is ob vious that a dieomorphism is b oth a smo oth map and a homeomorphism from the denitions. Ho w ev er, what is not so ob vious is that b y requiring a dieomorphism, w e are asking for more than something that is b oth smo oth and homeomorphism. Let us lo ok at some of follo wing coun terexamples and w e will learn what factors ma y con tribute to mak e the transformation fail to b e a dieomorphism. First, a smo oth bijection is not necessarily a dieomorphism. Let M 1 R b e [0 ; 2 ) and M 2 = S 1 the unit circle. f : M 1 M 2 dened b y f ( ) = (cos( ) ; sin ( )) is smo oth and bijectiv e, but not a dieomorphism b ecause the in v erse f 1 is not con tin uous, and hence not dieren tiable. In fact, f is not a homeomorphism. Second, a smo oth homeomorphism is not necessarily a dieomorphism. Consider f : R R with f ( x ) = x 3 It is smo oth and it is a homeomorphism but the in v erse f 1 is not dieren tiable at x = 0. A related concept is lo c al die omorphism and it is dened as: Denition 3. A dier entiable map f : M 1 M 2 is a lo c al die omorphism if for e ach x 2 M 1 ther e exists a neighb orho o d U of x such that f j U : U f ( U ) is a die omorphism.

PAGE 19

8 F urthermore, it is useful that w e list without pro of some kno wn facts in dieren tial top ology [ 9 8 50 10 31 ] ab out dieomorphism and lo cal dieomorphism. Theorem 1. A map f : M 1 M 2 is a lo c al die omorphism if and only if its tangent map is an isomorphism. With some simple facts in linear algebra, the ab o v e theorem can b e rewritten as the follo wing. Theorem 2. A map f : M 1 M 2 is a lo c al die omorphism if and only if its Jac obian of the tangent map is nowher e e qual to zer o. Theorem 3. A map f : M 1 M 2 is a die omorphism if and only if it is a bije ction and a lo c al die omorphism. With the help of these theorems w e can visualize more situations when a smo oth map fails to b e dieomorphism. One situation is when the tangen t map fails to b e an isomorphism ev erywhere. Namely the Jacobian is zero at some p oin ts. In that case, the smo oth map ev en fails to b e a lo cal dieomorphism. The other situation is that the smo oth map is a lo cal dieomorphism but not a global dieomorphism. The follo wing example demonstrates this. Consider f : R 2 R 2 with f ( x; y ) = ( e x cos y ; e x sin y ). The Jacobian is J = @ f x @ x @ f x @ y @ f y @ x @ f y @ y = e x cos y e x sin y e x sin y e x cos y = e 2 x 6 = 0 8 ( x; y ) 2 R 2 : So f is a lo cal dieomorphism. Ho w ev er, f is not a dieomorphism as can b e seen in Figure. 1{1 Notice that b ecause the function is p erio dic in y in the co domain, the image of the function has innitely man y sheets o v erlaid. Another related example is the self in tersection of the immersion of the Klein b ottle in R 3 as sho wn in Figure. 1{2 No w let us lo ok at an example of another smo oth transformation, namely the thin-plate spline (TPS) [ 70 ].

PAGE 20

9 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 -6 -4 -2 0 2 4 6 200 150 100 50 0 50 100 150 200 150 100 50 0 50 100 150 (a) (b) Figure 1{1: Grids in R 2 (a) b efore transformation and (b) after transformation Figure 1{2: Klein b ottle immersed in R 3 Image courtesy of John Sulliv an.(h ttp://torus.math.uiuc.edu/jms/Images/klein.h tml)

PAGE 21

10 (a) (b) Figure 1{3: (a) The template image and (b) The reference image 0 50 100 150 200 0 20 40 60 80 100 120 140 160 180 200 0 50 100 150 200 0 20 40 60 80 100 120 140 160 180 200 (a) (b) Figure 1{4: Space deformations (a) obtained with thin-plate spline in terp olation. The folding of space is illustrated b y the deformation of the grid lines. (b) The desired space transformation: a dieomorphism, whic h eliminates the space folding problem.

PAGE 22

11 Figure 1{3 a sho ws a template image. Figure 1{3 b is the reference image obtained b y w arping the template. Some landmarks are selected and sho wn in the images. Figure 1{4 a demonstrates the transformation of space b y sho wing the deformation of the rectangular grid. W e can see the folding of space. This is the dra wbac k of the thin-plate spline in terp olation. Due to the folding of space, features in the template ma y b e smeared in the o v erlapping regions. And furthermore, the transformation is not in v ertible. A dieomorphic transformation is strongly desirable, whic h preserv es the features, the top ology and whic h is smo oth as sho wn in Figure 1{4 b. W e still face the unkno wn corresp ondence problem, whic h is a dicult problem. There are t w o scenarios in whic h w e encoun ter the corresp ondence problem. The rst scenario is when w e do landmark based image registration. If the landmarks are selected b y hand, w e do kno w the corresp ondence. Ho w ev er, the pro cess of hand pic king landmarks is painstaking and it requires exp ert kno wledge ab out the image and the sub ject area and it ma y in v olv e h uman error. The complete automation of landmark selection is still not ac hiev ed at presen t but there is go o d progress to w ards the goal. If the landmarks are automatically selected, then the corresp ondence is unkno wn and the corresp ondence needs to b e automatically obtained. The second scenario is when w e ha v e a p oin t set represen tation of the shap es. Eac h shap e is represen ted b y a large p oin t set (p oin t cloud) and w e ha v e no kno wledge ab out the corresp ondence b et w een the t w o p oin t sets. What mak es the unkno wn corresp ondence problem more dicult is that it is ill-p osed. The corresp ondence b et w een t w o p oin t sets is a v ery in tuitiv e concept whic h ev ery one seems to understand. Ho w ev er, as is w ell kno wn within the medical image analysis comm unit y it is v ery dicult to dene corresp ondence precisely This creates a problem for v alidation since an irreducibly sub jectiv e factor seems to b e presen t in deciding what is a go o d corresp ondence.

PAGE 23

12 W e con tin ue with a short discussion aimed at reac hing a b etter understanding of p oin t corresp ondence. Supp ose w e ha v e t w o p oin t sets S 1 = f p i 2 n 1 j i = 1 ; 2 ; :::; n g and S 2 = f q i 2 n 2 j i = 1 ; 2 ; :::; n g where n 1 R d and n 2 R d usually with d = 2, or d = 3. When S 1 and S 2 are p oin t sets of equal cardinalit y with p oin ts randomly distributed in space, what is the corresp ondence b et w een S 1 and S 2 ? There are n bijections, or p erm utations b et w een the t w o sets. W e ha v e no a priori kno wledge allo wing us to judge that one corresp ondence is b etter than the other. Finite p oin t sets ha v e no extra structure o v er and b ey ond their discrete structure. So when w e talk ab out the corresp ondence b et w een p oin t sets, w e alw a ys imply that the p oin t set is the represen tation of some underlying shap e, whic h is a top ological space. And, as is fairly standard in the literature, w e can b ypass discussing p oin t corresp ondence b y fo cusing on the corresp ondence of the underlying top ological space, whic h is a homeomorphism, and most desirably a dieomorphism. Ev en so, the homeomorphism b et w een the t w o top ological spaces is not unique. Here w e mak e our rst assumption: the optimal corresp ondence b et w een the t w o p oin t sets is the one that induces the least deformation of space. Here w e understand the notion of space deformation in tuitiv ely but w e will carefully dene it later. It is easy to realize that this assumption is not alw a ys true b ecause w e are only giv en the p oin t set and w e ha v e no explicit kno wledge of the underlying shap e. This is easily illustrated with a simple example. In Figure 1{5 a and 1{5 b w e ha v e t w o p oin t sets. It is not clear at all what the corresp ondence is b et w een the t w o p oin t sets. Figure 1{5 c and 1{5 d put the t w o p oin t sets on top of t w o underlying images, the images of t w o w omen. Figure 1{5 c is the dra wing My Wife and My Mother-in-law published in 1915 b y the carto onist W. E. Hill. Figure 1{5 d is the image of He ad of a Woman in Pr ole (cropp ed) b y Jean-Pierre Da vid. If y ou in terpret Figure 1{5 c as the image of a y oung girl, namely the p oin t C as the

PAGE 24

13 50 100 150 200 50 100 150 200 250 A B C D E F G H 50 100 150 200 50 100 150 200 250 A B C D E F G H I J K L M (a) (b) A B C D E F G H A B C D E F G H I J K L M (c) (d) Figure 1{5: Am biguit y of landmark corresp ondence

PAGE 25

14 c hin and H as the ear, then w e ma y ha v e the corresp ondence b et w een the t w o p oin t sets from Figure 1{5 a to 1{5 b: A A, B B, C C, D D, E E, F F, G G and H H. F rom another p ersp ectiv e, if w e view the image as that of an old w oman, namely p oin t C as the tip of the nose, E the c hin and H the left ey e, w e ma y ha v e the corresp ondence: A A, B B, C I, D J, E C, F K, G L and H M. Let us lo ok at another example. Figure 1{6 a and Figure 1{6 b are t w o images of a blac k strip with the latter non-rigidly deformed. W e select four landmarks A1 through A4 for shap e A and four landmarks B1 through B4 for shap e B. Figure 1{6 c sho ws the landmarks for shap e A without the underlying ribb on shap e. Figure 1{6 d sho ws the landmarks for shap e B without the underlying ribb on shap e. Giv en the fact that w e can see the shap es, w e kno w the corresp ondences are A 1 B 1, A 2 B 2, A 3 B 3, A 4 B 4. Figure 1{7 a sho ws this landmark corresp ondence and Figure 1{7 b sho ws the deformation of space for this landmark corresp ondence. Ho w ev er, if w e only ha v e the t w o sets of p oin ts as sho wn in Figure 1{6 c and Figure 1{6 d, without the kno wledge of the underlying shap es, the ab o v e corresp ondence is not the one that w e will nd since it do es not giv e the least deformation of space. In fact, another corresp ondence, A 1 B 1, A 2 B 3, A 3 B 2, A 4 B 4 sho wn in Figure 1{7 d giv es a smaller deformation solution as sho wn in Figure 1{7 e. This is in fact a misin terpretation of the shap e, with Figure 1{7 c as the original shap e and Figure 1{7 f the misin terpreted shap e, due to the lac k of information regarding the underlying shap es. Consequen tly b y making the ab o v e assumption of least space deformation, w e are really assuming that the p oin ts are dense enough suc h that the underlying shap e is w ell represen ted b y the p oin ts. Adding more p oin ts can help resolv e this am biguit y and hence k eep the ab o v e assumption|that the correct corresp ondence giv es the smallest deformation of space|appro ximately v alid.

PAGE 26

15 A1 A2 A3 A4 B1 B2 B3 B4 (a) (b) 0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200 A1 A2 A3 A4 0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200 B1 B2 B3 B4 (c) (d) Figure 1{6: Tw o shap es of a ribb on and the asso ciated landmarks. (a) Shap e A, (b) Shap e B, (c) Landmarks of shap e A without the underlying shap e, (d) Landmarks of shap e B without the underlying shap e

PAGE 27

16 0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200 A1 B1 A4 B4 A2 B2 B3 A3 0 50 100 150 200 0 20 40 60 80 100 120 140 160 180 B1 B2 B3 B4 (a) (b) (c) 0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200 A1 B1 A4 B4 A2 B2 B3 A3 0 50 100 150 200 0 20 40 60 80 100 120 140 160 180 200 B1 B2 B3 B4 (d) (e) (f ) Figure 1{7: (a) The \correct" corresp ondence. (b) The space deformation according to the \correct" corresp ondence is not the smallest. (c) The \correctly" deformed shap e. (d) Another p ossible but \incorrect" corresp ondence. (e) The space deformation with this \incorrect" corresp ondence is the smallest. (f ) The misin terpretation of the shap e due to the \incorrect" corresp ondence. The rest of the dissertation is organized as follo ws. In Chapter 2 w e briery review the previous related w ork in the past ten y ears. In Chapter 3 w e describ e our theory of dieomorphic p oin t matc hing, dev elop an algorithm to solv e the problem and apply the algorithm to 2D and 3D medical imaging applications. Chapter 4 is an extension to the w ork in Chapter 3 and in it w e describ e a metho d of top ological clustering and matc hing. In Chapter 5 w e summarize our con tributions and p oin t out the directions of future w ork.

PAGE 28

CHAPTER 2 PREVIOUS W ORK In this c hapter, w e giv e a brief accoun t of the previous researc h that is related to our w ork. Most of the w ork w as dev elop ed in the past ten y ears but some w ork ma y b e traced bac k to the 1970s. W e start with thin-plate splines, whic h pla y an imp ortan t role in shap e matc hing, and all other splines w ork with the same principle. The Repro ducing Kernel Hilb ert Space form ulation giv es us a great insigh t on the nature of the en tire class of spline w arping problems. W e then discuss the folding problem with splines and v arious approac hes to o v ercome this, including imp osing constrain ts on the Jacobian and the ro w approac hes. W e also discuss the corresp ondence problem. In general there are t w o classes of approac hes to tac kle the corresp ondence problem. One is to confron t it directly while the other is to try to circum v en t it. With the second class of approac hes, p eople ha v e used implicit corresp ondence b y using a shap e distance that is the function of the t w o p oin t sets, instead of dep ending on the p oin t-to-p oin t corresp ondence, and there are also some metho ds that transform the p oin t matc hing problem to image matc hing problem through distance transform. With the rst class of approac hes, the direct approac h, there are metho ds that treat the corresp ondence and matc hing separately lik e softassign and there are metho ds that handles b oth at the same time lik e the join t clustering and matc hing (JCM) algorithms. Shap e con text metho d mak es use of extra lo cal con text information to help b etter resolv e the am biguous p oin t corresp ondence problem. Activ e shap e mo dels pla y as a bridge b et w een the landmark based metho ds and image in tensit y based metho ds, and also a bridge b et w een rigid matc hing and non-rigid matc hing. W e sp end a section discussing the deterministic annealing metho d for EM clustering, whic h is an 17

PAGE 29

18 eectiv e metho d to a v oid lo cal minima and w e adopt suc h an annealing approac h in our algorithm. Also closely related is the w ork of statistical shap e analysis on dieren tiable manifold and some distance measures from Fisher information theory 2.1 Thin Plate Splines (TPS) A spline is a long strip of w o o d or metal that is xed at a n um b er of p oin ts. This has long b een used as a drafting to ol to dra w smo oth curv es that are required to pass certain p oin ts (called \duc ks," or \dogs," or \rats"). These w o o den strips are not only used in drafting, but also used in constructions. The W righ t brothers used one to shap e their wings. In the old da ys splines w ere used in shipbuilding. They are also used for b ending the w o o d for m usical instrumen ts lik e pianos, violins, violas, etc. In the mo dern da ys, splines are used to mo del the b o dy of automobiles. In all those uses, it is actually the ph ysical realization of some smo oth curv es. In ph ysics, the shap e of the w o o d strip has to tak e the form of suc h a curv e that the b ending energy of the strip is minimized. Sheo en b erg [ 63 62 ] is credited to b e the rst to study spline functions started with one dimensional problems, the cubic splines. The thin plate spline is the natural generalization of the cubic spline. In the drafting practice, the w o o den strip is long but v ery thin and narro w, so that it is abstracted as a one dimensional curv e. In the thin plate case, w e ha v e a plate whic h is v ery thin, so the problem is t w o dimensional. In b oth cases, the same ph ysical b ending energy of the strip or the thin plate in minimized. The problem of thin plate spline in terp olation is stated as follo ws. Giv en f p i 2 R 2 j i = 1 ; 2 ; :::; n g and f z i 2 R j i = 1 ; 2 ; :::; n g w e w an t to nd a function f : R 2 R 2 suc h that it in terp olates the p oin ts with f ( p i ) = z i i = 1 ; 2 ; :::n and minimizes the thin plate b ending energy E ( f ) = Z Z R 2 ( @ 2 f @ x 2 ) 2 + 2( @ 2 f @ x@ y ) 2 + ( @ 2 f @ y 2 ) 2 dxdy : (2.1)

PAGE 30

19 If w e in terpret the f z i 2 R j i = 1 ; 2 ; :::; n g as the displacemen ts of the thin plate in the z direction at p oin ts f p i 2 R 2 j i = 1 ; 2 ; :::; n g in the x y plane, w e can easily see the ph ysical in tuition of this in terp olation problem, as sho wn in Figure. 2{1 Sometimes it is desirable that w e sacrice the exact in terp olation. W e only seek an appro ximation of the in terp olation as a trade o of less deformation of the thin plate. So the thin plate smo othing problem is dened as giv en f p i 2 R 2 j i = 1 ; 2 ; :::; n g and f z i 2 R j i = 1 ; 2 ; :::; n g w e w an t to nd a function f : R 2 R 2 suc h that it minimizes the cost function 1 n n X i =1 ( z i f ( p i )) 2 + E ( f ) : (2.2) 2.2 Bo okstein's Application to 2D Landmark W arping Bo okstein [ 4 5 ] applied thin plate splines to the landmark in terp olation problem. F or simplicit y w e discuss the problem in 2D space. Ev erything in the 2D form ulation easily applies to 3D except w e ha v e a dieren t k ernel for 3D. The goal is to nd a smo oth transformation f : n 1 n 2 that in terp olates n pairs of landmarks f p i 2 n 1 j i = 1 ; 2 ; :::; n g and f q i 2 n 2 j i = 1 ; 2 ; :::; n g It is ob vious that suc h a smo oth transformation is not unique. W e will giv e it further constrain ts. One c hoice is to require the to transformation minimize some functional, in this case E = 2 X h =1 Z Z R 2 ( @ 2 f h @ x 2 ) 2 + 2( @ 2 f h @ x@ y ) 2 + ( @ 2 f h @ y 2 ) 2 dxdy ; (2.3) where f 1 and f 2 are the x and y comp onen ts of the mapping, f ( x; y ) = ( f 1 ( x; y ) ; f 2 ( x; y )) : (2.4) If w e in terpret eac h of f 1 and f 2 as the b ending in the z direction of a metal sheet, or thin plate, extending in the x y plane, the energy in ( 2.3 ) is the analogy of the

PAGE 31

20 thin plate b ending energy The problem can b e solv ed using a standard Green's function metho d. The Green's function for this problem is the function satisfying the equation 2 U ( x ; x 0 ) = ( @ 2 @ x 2 + @ 2 @ y 2 )( @ 2 @ x 2 + @ 2 @ y 2 ) U = ( x ; x 0 ) (2.5) where ( x ; x 0 ) is the Dirac delta function and x = ( x; y ) and x 0 = ( x 0 ; y 0 ) 2 R 2 This can b e solv ed with the solution U ( r ) = r 2 l og r 2 ; (2.6) where r is the distance p x 2 + y 2 Because an ane transformation has a zero con tribution to the b ending energy the transformation allo ws a free ane transformation. Dene the matrices K = 266666664 0 U ( r 12 ) ::: U ( r 1 n ) U ( r 21 ) 0 ::: U ( r 2 n ) ::: ::: ::: ::: U ( r n 1 ) U ( r n 2 ) ::: 0 377777775 ; whic h is n n ; (2.7) P = 266666664 1 x 1 y 1 1 x 2 y 2 ::: ::: ::: 1 x n y n 377777775 ; whic h is 3 n ; (2.8) and L = 264 K P P T O 375 ; whic h is ( n + 3) ( n + 3) ; (2.9) where the sym b ol T is the matrix transp ose op erator and O is a 3 3 matrix of zeros.

PAGE 32

21 Let V = ( v 1 ; :::; v n ) b e an y n -v ector and write Y = ( V j 0 0 0) T Dene the v ector W = ( w 1 ; :::; w n ) and the co ecien ts ( a 1 ; a x ; a y ) b y the equation L 1 Y = ( W j a 1 a x a y ) T : (2.10) Finally w e ha v e the solution f ( x; y ) = a 1 + a x x + a y y + n X i =1 w i U ( j P i ( x; y ) j ) : (2.11) 2.3 More on Splines Bo okstein's seminal w ork applying thin-plate splines to the landmark problems in 2D [ 4 5 ] is w ell kno wn in the computer vision and imaging comm unit y What is less kno wn is the theoretical foundations for the thin-plate spline laid b y Duc hon [ 21 22 23 ] and Meinguet [ 48 ] in the 1970s. This solution is correct and simple. Ho w ev er it is not n umerically stable b ecause it in v olv es the in v erse of the large k ernel matrices. A n umerically stable solution is giv en b y W ah ba [ 70 ] using the QR decomp osition. W ah ba has an accoun t for a more generalized form ulation for d -dimensional space and the energy in v olving m th order partial deriv ativ es. Here w e only discuss the sp ecial thin plate spline with d = 2 and m = 2. Later in our w ork in 3D situations, 3D thin plate splines are used where d = 3 but it is straigh tforw ard to generalize from 2D to 3D, while the k ernel in 3D is dieren t than that for 2D. Let f ; g 2 W k ; 2 (n), where W k ; 2 (n) is the Sob olev space. The inner pro duct of f and g is dened as A thin-plate smo othing problem in 2D is to nd f 2 W k ; 2 (n) to minimize the functional 1 n n X i =1 jj q i f ( p i ) jj 2 + E ( f ) : (2.12)

PAGE 33

22 The n ull space is spanned b y 1 ( t ) = 1 ; 2 ( t ) = x; 3 ( t ) = y ; (2.13) whic h is an ane space. Duc hon [ 23 ] pro v ed that if f p i 2 n 1 j i = 1 ; 2 ; :::; n g are suc h that least squares regression on 1 ; 2 ; 3 is unique, then 2.12 has a unique minimizer ^ f with the represen tation ^ f ( t ) = 3 X =1 d ( t ) + n X i =1 c i K ( t; p i ) ; (2.14) where K is a Green's function for the double iterated Laplacian. K has the form K = K ( j s t j ) ; (2.15) and K ( ) = 2 log j j : (2.16) By using in tegration b y part, w e obtain that c; d are the minimizers of 1 n jj y T d K c jj 2 + c 0 K c (2.17) sub ject to T 0 c = 0, where K is a n n matrix with ij th en tries K ( p i ; p j ) : Let us do a QR decomp osition on T T = ( Q 1 : Q 2 ) 0B@ R 0 1CA (2.18) where ( Q 1 ; Q 2 ) is orthogonal and R is lo w er triangular. Q 1 is n 3 and Q 2 is n ( n 3). Since T 0 c = 0, c m ust b e in the column space of Q 2 with c = Q 2 r for some n 3 v ector r Substituting in 2.17 and the energy is in the form of

PAGE 34

23 1 n jj Q 02 y Q 02 K Q 2 r jj 2 + 1 n jj Q 01 y R d Q 01 K Q 2 r jj 2 + r 0 Q 02 K Q 2 r : (2.19) The solutions are d = R 1 ( y K Q 2 r ) c = Q 2 r ; (2.20) with r = ( Q 02 K Q 2 + nI ) 1 Q 02 y (2.21) where I is the iden tit y matrix. 2.4 Repro ducing Kernel Hilb ert Space (RKHS) F orm ulation It is p ossible to come to the same result for the minimizer using repro ducing k ernels and from this p oin t of view it is easier to understand the existence and uniqueness of the solution [ 70 7 ]. Let jj f jj 2 = E = R n ( f 2 xx + 2 f 2 xy + f 2 y y ) dxdy where jj f jj is the norm of f in W k ; 2 (n). Since W k ; 2 (n) is a Hilb ert space, from Riesz represen tation theorem, for an y x 2 n, the ev aluation linear functional x : W k ; 2 (n) R ; x ( f ) = f ( x ) (2.22) has a represen ter u x 2 W k ; 2 (n) suc h that f ( x ) = < u x ; f > : (2.23) No w the problems is transformed to a new problem: nd a function f 2 W k ; 2 (n) with minimal norm jj f jj sub ject to constrain ts

PAGE 35

24 < u X i ; f > = v i ; i = 1 ; 2 ; :::; n : (2.24) F or X a ; X b 2 n, u ( X a ; X b ) = u X a ( X b ) is the k ernel of the repro ducing k ernel Hilb ert space. Let T b e the linear subspace spanned b y u x i ; i = 1 ; 2 ; :::; n An y function f 2 W k ; 2 (n) can b e decomp osed in to f = f T + f ? where f T 2 T and f ? is in the orthogonal complemen t of T and hence < u X i ; f ? > = 0. W e kno w if f T satises ( 2.24 ), then f also satises ( 2.24 ) only with jj f jj > jj f T jj if f ? 6 = 0. So w e only need to searc h for the solution in T The general solution th us can b e written as f ( X ) = a 0 + a 1 x + a 2 y + n X i =1 w i u ( X i ; X ) ; where functions of the form a 0 + a 1 x + a 2 y span the n ull space. With this form, E can b e rewritten as n X i =1 ;j =1 w i U ij w j = W U W + : (2.25) 2.5 The F olding Problem of Splines This is all v ery nice except as men tioned ab o v e there is no mec hanism to guaran tee a bijection in order to b e homeomorphic or dieomorphic. In tuitiv ely this problem is kno wn as the folding of space. Figure 2{2 sho ws the displacemen t of landmarks. Figure 2{3 is the thin-plate spline in terp olation. W e can see the folding of space. This is the dra wbac k of thinplate spline in terp olation. Due to the folding of space, features in the template ma y b e smeared in the o v erlapping regions. And furthermore, the transformation is not in v ertible. A dieomorphic transformation is strongly desirable, since it preserv es the features, the top ology and is smo oth as sho wn in Figure 2{4

PAGE 36

25 2.6 Imp osing Restriction on the Jacobian One straigh tforw ard approac h to nd a dieomorphism for practical use is to mak e some remedy on the thin-plate spline. W e can restrict our searc h space to the set of dieomorphisms and the ideal one should minimize the thin-plate energy W e mak e the observ ation that if the Jacobian of the transformation f c hanges sign at a p oin t, then there is folding. W e put the constrain t requiring the Jacobian is alw a ys p ositiv e. There is some literature on this approac h but most of these approac hes do es not guaran tee that the transformation is smo oth ev erywhere [ 38 11 ]. 2.7 The Flo w Approac h Another approac h is to utilize a ro w eld[ 28 39 49 ]. W e in tro duce one parameter, the time t in to the dieomorphism. Let t : n n b e the dieomorphism from n to n at time t A p oin t x is mapp ed to the p oin t t ( x ). Sometimes w e also denote this as ( x; t ). It is easy to v erify that for all the v alues of t t form a one parameter dieomorphism group. If x is xed, then ( x; t ) traces a smo oth tra jectory in n. The in terp olation problem b ecomes: nd the one parameter dieomorphic group ( ; t ) : n n suc h that giv en p i 2 n and q i 2 n 8 i = 1 ; 2 ; :::; n ( x; 0) = x and ( p i ; 1) = q i W e in tro duce the v elo cit y eld v ( x; t ) and construct a dynamical system using the transp ort equation @ ( x; t ) @ t = v ( ( x; t ) ; t ) : (2.26) The in tegral form of the relation b et w een ( x; t ) and v ( x; t ) is ( x; 1) = x + Z 1 0 v ( ( x; t ) ; t ) dt : (2.27) Ob viously suc h a ( x; t ) is not unique and there are innitely man y suc h solutions. With the analogy to the thin-plate spline, it is natural that w e require the desirable dieomorphism results in minimal space deformation. Namely w e require the deformation energy

PAGE 37

26 Z 1 0 Z n jj Lv ( x; t ) jj 2 dxdt (2.28) to b e minimized, where L is a giv en linear dieren tial op erator. The follo wing theorem [ 39 ] states the existence of suc h a v elo cit y eld and sho ws a w a y to solv e for it. Theorem 4. L et p i 2 n and q i 2 n 8 i = 1 ; 2 ; :::; n The solution to the ener gy minimization pr oblem ^ v ( ) = arg min Z 1 0 Z n jj Lv ( x; t ) jj 2 dxdt (2.29) subje ct to ( p i ; 1) = q i ; 8 i = 1 ; 2 ; :::; n (2.30) wher e ( x; 1) = x + Z 1 0 v ( ( x; t ) ; t ) dt (2.31) exists and denes a die omorphism ( ; 1) : n n The optimum velo city eld ^ v and the die omorphism ^ ar e given by ^ v ( x; t ) = n X i =1 K ( ( x i t ) ; x ) n X j =1 ( K ( ( t )) 1 ) ij ^ ( x j ; t ) (2.32) wher e K ( ( t )) = 0BBBBBBBBBB@ K ( ( p 1 ; t ) ; ( p 1 ; t )) K ( ( p 1 ; t ) ; ( p n ; t )) K ( ( p n ; t ) ; ( p 1 ; t )) K ( ( p n ; t ) ; ( p n ; t )) 1CCCCCCCCCCA (2.33)

PAGE 38

27 with ( K (( ( t )) ij denoting the ij 3 3 blo ck entry ( K ( ( t )) ij = K ( ( p i ; t ) ; ( p j ; t )) and ^ ( p n ; ) = arg min ( p n ; ) Z 1 0 X ij ( p i ; t ) T ( K ( ( t )) 1 ) ij ( p j ; t ) dt (2.34) subje ct to ( p i ; 1) = q i ; i = 1 ; 2 ; :::; N with the optimal die omorphism given by ^ ( x; 1) = x + Z 1 0 ^ v ( ^ ( x; t ) ; t ) dt : (2.35) The pro of [ 39 ] is omitted here. With this theorem, w e can con v ert the original optimization problem lo oking for the v ector eld ^ v ( x; t ) to a problem of nite dimensional optimal con trol with end p oin t conditions. This problem is called the exact matching pr oblem b ecause w e required the images of the giv en p oin ts p i ; i = 1 ; 2 ; :::; n are exactly another set of giv en p oin ts q i ; i = 1 ; 2 ; :::; n The exact matc hing problem is symmetric with resp ect to t w o sets of landmarks or t w o p oin t shap es. Sw apping the t w o p oin t sets f p i 2 n 1 j i = 1 ; 2 ; :::; n g and f q i 2 n 2 j i = 1 ; 2 ; :::; n g results in the new optimal dieomorphism to b e the in v erse of the old dieomorphism. The exact matc hing problem can b e generalized to the inexact matching pr oblem. In the inexact matc hing problem, w e do not require the p oin ts exactly matc h. Instead, w e seek a compromise b et w een the closeness of the matc hing p oin ts and the deformation of space. W e minimize Z 1 0 Z n jj Lv ( x; t ) jj 2 dxdt + n X i =1 jj q i ( p i ; 1) jj 2 ; (2.36) whic h can b e similarly solv ed. As seen from the form ulation of the problem, there ma y b e innitely man y dieomorphisms that in terp olate the t w o sets of landmarks. A usual w a y to nd a particular desirable dieomorphism is to require the dieomorphism to minimize a

PAGE 39

28 certain ob jectiv e function. Camion and Y ounes [ 7 ] prop osed a dieren t ob jectiv e function in the form of E ( v ; q ) = Z 1 0 Z n jj Lv ( x; t ) jj 2 dxdt + n X i =1 jj dq i ( t ) dt v ( q i ( t ) ; t ) jj 2 dt (2.37) o v er all time dep enden t v elo cities v ( x; t ) on n and o v er all tra jectories q 1 ( t ) ; :::; q n ( t ). This can b e in terpreted as a geo desic distance b et w een t w o p oin ts on the conguration manifold. This is generalized to exact matc hing b y Marsland et al. [ 46 ]. 2.8 Corresp ondence and Softassign So far all w ork has assumed landmark matc hing with a kno wn corresp ondence. When the corresp ondence is unkno wn, the problem is dramatically complicated. One approac h is the softassign [ 54 43 ] metho d, whic h solv es for the corresp ondence as a p erm utation problem using linear programming in the space of doubly sto c hastic matrices [ 65 3 ]. The problem of nding the corresp ondence b et w een t w o sets of p oin ts can b e form ulated as nding the p erm utation matrix. This approac h is exp ensiv e in computational time. Ch ui et al. [ 12 17 14 13 15 ] adopted the join t clustering sc heme, in whic h the corresp ondence and space deformation is estimated sim ultaneously Ho w ev er, in all the w ork, splines are used and a p oten tial dra wbac k is that a dieomorphic mapping in space is not guaran teed. 2.9 Distance T ransforms An un usual w a y of getting around the p oin t corresp ondence problem is to use distance transforms to con v ert the p oin t matc hing problem to an image matc hing problem. The distance transform w as rst in tro duced b y Rosenfeld and Pfaltz [ 60 ], and it has a wide range of applications in image pro cessing, rob otics, pattern recognition and pattern matc hing. P aragios et al. [ 53 ] giv e one example of using distance transforms to establish lo cal corresp ondences for compact represen tations

PAGE 40

29 of anatomical structures. Distance transform applies to binary images, as w ell as p oin t sets, whic h can b e though t as a sp ecial case of a binary image. Supp ose w e ha v e a domain n, and a p oin t set S n. F or eac h p oin t x 2 n, w e assign to it a non-negativ e real n um b er, whic h is the shortest distance from x to all the p oin ts in S This w a y w e obtain a scalar eld in domain n. W e treat this scalar eld as a gra y scale image, w e call it the distance transformed image of the p oin t set. If w e ha v e t w o p oin t sets, w e can rst p erform the distance transform on the t w o sets resp ectiv ely and later register the t w o distance transformed images. Ho w ev er, there are problems with this approac h. If the matc hing of the t w o distance transformed images is in tensit y based, then the original p oin ts ma y not b e matc hed exactly Ev en in the inexact matc hing case, the optimization for the distance transformed images is o v er the en tire image region and that ma y not b e optimal for the original p oin t set. If the matc hing of the t w o distance transformed images is lev el set based, the lev el sets in the t w o images ma y not b e top ologically equiv alen t, as sho wn in Figure 2{5 and Figure 2{6 In Figure 2{5 there are t w o distance transformed images, with three p oin ts eac h in the image. Figure 2{6 sho ws the lev el sets of the t w o distance transformed images. Hence this metho d cannot guaran tee to obtain a dieomorphism. Due to the indirect approac h of transforming p oin t sets in to distance transforms, this metho d has not seen wide applicabilit y for p oin t sets. 2.10 Implicit Corresp ondence It is p ossible to dene distance measures b et w een t w o shap es, with the shap es view ed as p oin t sets, without the kno wledge of corresp ondence. Hausdor distance b et w een t w o p oin t sets is a w ell-kno wn distance measure for suc h purp oses. Huttenlo c her et al. [ 37 ] dev elop metho ds comparing shap es using Hausdor distance. Glaunes et al. [ 27 ] use another metho d to circum v en t the corresp ondence problem b y in tro ducing a distance measure of the t w o shap es using an external

PAGE 41

30 Hilb ert space structure. If the t w o shap es are p oin t sets, the distance measure b et w een t w o p oin t sets used in Glaunes et al. [ 27 ] is 1 2 R j v1 v j 2I (2.38) where j j 2I is the norm squared in some Hilb ert space I and j j 2I = N 1 X i =1 N 2 X j =1 c i c j k I ( x i ; y j ) (2.39) where c i ; c j are constan t co ecien t. x i are the p oin ts in the rst sets and y j are the p oin ts in the second sets. k I is some k ernel. In their exp erimen ts, they used radial basis function k ernels k I ( x; y ) = f I ( j x y j 2 2 I ) ; (2.40) with f I ( u ) = e u and f I ( u ) = 1 1+ u The idea of distance measure without explicit corresp ondence actually go es bac k to Grimson et al. [ 29 ] and Lu and Mjolsness [ 45 ] in 1994. The distance measure used is X i X j e jT l i m j j 2 2 2 : (2.41) As p oin ted out in Rangara jan et al. [ 55 ] this is equiv alen t to using E ( M ; T ) = X ij M ij D ij ( T ) + X ij M ij log M ij ; (2.42) where M ij is the explicit corresp ondence and T is the space transformation and D ij is the distance measure b et w een p oin t pairs that are in corresp ondence. Guo et al. suggested a join t clustering algorithm for solving 2D dieomorphic p oin t matc hing problems, with unkno wn corresp ondence, using explicit corresp ondences [ 33 32 ]. The dieren t form ulations with explicit corresp ondence and implicit corresp ondence

PAGE 42

31 are closely related to eac h other, through a Legendre transform, as p oin ted out b y Mjolsness and Garrett [ 51 ]. 2.11 Shap e Con text In the approac h of shap e con text [ 2 ], the corresp ondence is expressed explicitly and the corresp ondence problems is tac kled directly Moreo v er, the corresp ondence is solv ed separately from the space transformation with the help of shap e c ontext. Shap e con text is the lo cal shap e information at eac h p oin t. F or eac h p oin t q and all other p oin ts p i w e can dra w a v ector from q to p i that is p i q The lo cal shap e con text information is all stored in this set of n 1 v ectors. This information is ric h and in practice the distribution of these n 1 v ectors giv es us more robust, compact and discriminativ e descriptor. F or eac h p oin t p i w e compute a histogram h i of these n 1 v ectors h i ( k ) = # f q 6 = p i j ( q p i ) 2 bin( k ) g : (2.43) The histogram is dened to b e the shap e con text of p oin t p i Bins that are uniform in log-p olar space are used to mak e the nearb y con text p oin ts more imp ortan t the the far a w a y con text p oin ts. The cost for the corresp ondence b et w een t w o shap es is the sum of the cost of corresp onding p oin t pairs H ( ) = X i C ( p i ; q ( i ) ) ; (2.44) whic h is a function of the p erm utation and the cost with eac h individual pair of p oin ts is dened as C ij = C ( p i ; q j ) = 1 2 K X k =1 [ h i ( k ) h j ( k )] 2 h i ( k ) + h j ( k ) : (2.45) 2.12 Activ e Shap e Mo dels Co otes et al. prop ose a metho d, whic h they call Activ e Shap e Mo dels [ 18 ] to lo cate ob jects in images, with the help of mo del shap es and the training of

PAGE 43

32 these mo del shap es. This approac h pla ys as a bridge b et w een the landmark based metho ds and image in tensit y based metho ds, and also a bridge b et w een rigid matc hing and non-rigid matc hing. They use landmark p oin ts on the mo dels but not on the test images. The landmarks are hand pic k ed and hand lab eled. The automatic corresp ondence problem is circum v en ted b y the h uman in v olv emen t of the landmark pic king and lab eling pro cess. The deformation mo del of the template is similarit y whic h is v ery close to rigid, allo wing translation, rotation plus a scaling. What mak es it applicable to non-rigid deformation is that they learn the statistics of the mo del shap es through a set of training samples and nd the mean and v ariance in higher dimensional space and apply the PCA analysis with the v ariances. When it applies to lo cating the mo del in the image, they use the snak e mo del, whic h is non-rigid in nature, with the help and restriction of the statistics of the mo del shap es. 2.13 Deterministic Annealing Applied to EM Clustering Deterministic annealing is an eectiv e tec hnique used in the clustering problems. The clustering problem is a non-con v ex optimization problem. The traditional clustering tec hniques use descen t based algorithms and they tend to get trapp ed in a lo cal minim um. Rose et al. [ 59 ] prop osed an annealing approac h using analogies to statistical ph ysics. The clustering problem is to partition a set of data p oin ts, f x i 2 R d j i = 1 ; 2 ; :::; n g in to K clusters C 1 ; C 2 ; :::; C K with the cen ters r 1 ; r 2 ; :::; r K resp ectiv ely In the fuzzy clustering literature, w e call the probabilit y of eac h p oin t p i b elonging to eac h cluster C j the fuzzy mem b ership. Hard clustering is a marginal sp ecial case, where eac h p oin t is deterministically asso ciated with a single cluster. Let E j ( x i ) denote the energy asso ciated with assigning a data p oin t x i to cluster C j

PAGE 44

33 The a v erage total energy is < E > = X i X j P ( x i 2 C j ) E j ( x i ) : (2.46) Since w e do not mak e an y assumption ab out the data distribution, w e apply the principle of maxim um en trop y It is w ell kno wn from statistical ph ysics, the asso ciation probabilities P ( x i 2 C j ) that maximize the en trop y under constrain t ( 2.46 ) are Gibbs canonical distributions, P ( x i 2 C j ) = 1 Z i e E j ( x i ) ; (2.47) where Z i is the partition function Z i = X k e E k ( x i ) : (2.48) The parameter is the Lagrange m ultiplier determined b y the giv en v alue of < E > in ( 2.46 ). In the analogy in statistical ph ysics, is in v ersely prop ortional to the temp erature T W e ha v e assumed that w e ha v e a xed set of clusters. W e w an t to extend this to include optimization o v er the n um b er of clusters as w ell. The optimal solution then is the one that minimizes the free energy F = 1 ln Z : (2.49) The set of cluster cen ters are the ones that satisfy @ F @ r j = 0 ; 8 j : (2.50) And the solution is r j = P i x i P ( x i 2 C j ) P i P ( x i 2 C j ) : (2.51) This pro cedure determines a set of cluster cen ters f r j 2 R d j j = 1 ; 2 ; :::; n g for eac h xed Generally c hanging K the imp osed n um b er of clusters, will mo dify the p ositions of the set of cluster cen ters. Ho w ev er, there exists some n c suc h that

PAGE 45

34 for all K > n c one gets only n c distinct cluster cen ters while the remaining K n c cluster cen ters are rep etitions from this set. Th us at eac h giv en w e get at most n c clusters. Here w e assume K n c at a giv en and w e only consider them without rep etitions. The free energy F and are Legendre transform images of eac h other. Fixing one of them determines the other. F or = 0, eac h data p oin t is uniformly asso ciated with all clusters and all the cen ters ha v e the same lo cation, the cen troid of the data. Clearly for = 0 w e ha v e a single minim um, whic h is the global minim um, for F and the en tire data set in in terpreted as one cluster. A t higher the free energy ma y ha v e man y lo cal minima, and the concept of annealing emerges here can b e view ed as trac king the global minim um while gradually increasing Moreo v er, at = 0 there is only one cluster ( n c = 1), but at some p ositiv e w e shall ha v e n c > 1. In other w ords, this cluster will split in to smaller clusters, and will th us undergo a phase tr ansition The rst phase transition o ccurs at a critical v alue for c = 1 2 max ; (2.52) where max is the largest eigen v alue of C xx the co v ariance matrix of the data set. Giv en = 1 =T and for Gaussian mixture clustering T = 2 2 T it is quite understandable that the critical v alue c for the prescrib ed T is c = p max 2.14 Statistical Shap e Analysis on Dieren tiable Manifolds Da vid Kendall rst in tro duced the idea of represen ting shap es in complex projectiv e spaces. The idea is dev elop ed to represen t shap es on general dieren tiable manifolds, or shap e spaces [ 66 ]. A shap e is constructed from a sequence of landmarks p 1 ; p 2 ; :::p n with p i 2 R 2 Namely eac h landmark is a p oin t in 2-dimensional Euclidean space. The t w o sequences of landmarks in R 2 are considered of the same shap e if they dier only b y a similarit y transformation in R 2 A shap e then is

PAGE 46

35 dened as an equiv alen t class of these landmark sequences and the shap e space is carv ed out of the quotien t space dened b y the equiv alen t relation. One w a y to distill the shap e information out of the landmark sequence is to remo v e the lo cation, scale and orien tation. T o remo v e the lo cation, w e can dene r i = x i x ; (2.53) where x = 1 n n X j =1 x j (2.54) is the mean or cen troid of the landmarks. That is, to remo v e the lo cation, w e mak e the landmarks mean zero. T o remo v e the scale w e can mak e the v ariance of the landmarks as one. So w e dene i = x i x q P nj =1 jj x j x jj 2 : (2.55) W e refer the v ector as the pr e-shap e of the landmarks. The pre-shap e space is the in tersection of the ( n 2)-dimensional subspace F 2 n 2 = f ( x 1 ; :::; x n ) 2 R 2 n j n X j =1 x j = 0 g (2.56) with the unit sphere S 2 n 1 = f ( x 1 ; :::; x n ) 2 R 2 n j n X j =1 jj x j jj 2 = 1 g : (2.57) The in tersection S 2 n 3 = F 2 n 2 \ S 2 n 1 (2.58) is a (2 n 3)-dimensional sphere within the am bien t Euclidean space R 2 n It is more dicult to remo v e the orien tation of the shap e. T o do this, w e dene the orbit of a pre-shap e 2 S 2 n 3 is the circle O ( ) = f ( ) j 0 < 2 g S 2 n 3 : (2.59)

PAGE 47

36 Tw o pre-shap es are of the same shap e if they are on the same orbit. A shap e is dened as the equiv alen t class of the pre-shap es. If 1 and 2 are t w o preshap es, then the great circle distance b et w een 1 and 2 is giv en b y d ( 1 ; 2 ) = cos 1 ( < 1 ; 2 > ) : (2.60) The induced metric on n2 is then dened as d [ O ( 1 ) ; O ( 2 )] = inf f d ([ 1 ( 1 ) ; 2 ( 2 )] j 0 1 ; 2 < 2 g : (2.61) This is called Pr o cruste an distanc e. It can b e pro v ed that if 1 and 2 are t w o represen tativ es of t w o shap es, the pro crustean distance b et w een the t w o shap es can b e expressed as d ( 1 ; 2 ) = cos 1 ( j n X k =1 1 k 2 k j ) : (2.62) The standard statistics lik e means and v ariances can b e p erformed on the shap e manifold. One dra wbac k of this approac h is that the shap e is treated as the se quenc e of landmarks instead of sets. That means if w e t w o iden tical sets of landmarks and only lab el them dieren tly this theory treats them as t w o distinct shap e. It do es not consider to mak e an equiv alen t class from the p erm utation of landmark p oin ts. 2.15 Distance Measures from Information Theory Information geometry is an emerging discipline that studies the probabilit y and information b y w a y of dieren tial geometry In information geometry every probabilit y distribution is a p oin t in some space. A family of distributions corresp onds to p oin ts on a dieren tiable manifold. Endres et al. [ 25 ] prop osed a metric for t w o distributions. Giv en t w o probabilit y distributions P and Q and R = 1 2 ( P + Q ), w e can dene a distance D P Q

PAGE 48

37 b et w een the t w o distributions b y D 2 P Q = 2 H ( R ) H ( P ) H ( Q ) = D ( P jj R ) + D ( Q jj R ) = N X i =1 ( p i log 2 p i p i + q i + q i log 2 p i p i + q i ) : (2.63) This metric can also b e in terpreted as the square ro ot of an en trop y appro ximation to the logarithm of an evidence ratio when testing if t w o samples ha v e b een dra wn from the same underlying distribution. 1 2 D 2 P Q is named Jensen-Shannon div ergence, whic h is dened as D ( P ; Q ) = D ( P jj R ) + (1 ) D ( Q jj R ) (2.64) R = P + (1 ) Q and therefore 1 2 D 2 P Q = D 1 2 ( P ; Q ) : (2.65) With probabilit y distribution P ( x j ) where is a set of parameters 1 ; :::; n the Fisher information is dened as [ 26 ] G ij ( ) = E @ 2 log p ( x j ) @ i @ j = Z p ( x j ) @ 2 log p ( x j ) @ i @ j dx : (2.66) C. R. Rao [ 57 ] suggested this is a metric. In fact, it is the only suitable metric in parametric statistics and it is called Fisher-Rao metric.

PAGE 49

38 5 10 15 20 5 10 15 20 0.2 0 0.2 0.4 Figure 2{1: Deformation of the thin plate 0 20 40 60 80 100 0 10 20 30 40 50 60 70 80 90 100 Figure 2{2: Landmark displacemen ts

PAGE 50

39 0 20 40 60 80 100 120 0 20 40 60 80 100 120 140 160 Figure 2{3: Thin-plate Spline in terp olation 20 0 20 40 60 80 100 120 140 20 0 20 40 60 80 100 120 140 160 Figure 2{4: Dieomorphic in terp olation

PAGE 51

40 0 100 200 300 400 500 600 100 200 300 400 500 600 (a) 0 100 200 300 400 500 600 100 200 300 400 500 600 (b) Figure 2{5: Tw o distance transformed images of three landmarks

PAGE 52

41 0 100 200 300 400 500 600 100 200 300 400 500 600 (a) 0 100 200 300 400 500 600 100 200 300 400 500 600 (b) Figure 2{6: Lev el sets of t w o distance transformed images

PAGE 53

CHAPTER 3 DIFFEOMORPHIC POINT MA TCHING In this c hapter w e in v estigate the dieomorphic p oin t matc hing theory and apply the theory to shap es in medical imaging. In Section 3.1 w e pro v e a theorem ab out the existence of a dieomorphic mapping matc hing the landmarks. In Section 3.2 w e pro v e another theorem ab out the symmetric nature in the case of exact landmark matc hing. In Section 3.3 w e form ulate a theory of dieomorphic p oin t matc hing with unkno wn corresp ondence and devise an ob jectiv e function. In Section 3.4 w e design an algorithm to solv e the problem using join t clustering and deterministic annealing. In Section 3.5 w e apply the algorithm to 2D corpus callosum shap es. In Section 3.6 p erform matc hing on 3D hipp o campus shap es. Both corpus callosum and hipp o campus are parts in the h uman brain and the matc hing of these shap e ha v e great signicance in medical treatmen t and medical researc h. 3.1 Existence of a Dieomorphic Mapping in Landmark Matc hing W e ha v e discussed in Chapter 1 that in 2D a similarit y transformation can map exactly 2 giv en p oin ts to 2 giv en p oin ts. A ane transformation can map exactly 3 giv en p oin ts to 3 giv en p oin ts. A pro jectiv e transformation can map exactly 4 giv en p oin ts to 4 giv en p oin ts. No w giv en n arbitrary distinct p oin ts f p i 2 R 2 j i = 1 ; 2 ; :::n g and another set of n arbitrary distinct p oin ts f q i 2 R 2 j i = 1 ; 2 ; :::n g w e w an t to nd a dieomorphism f : R 2 R 2 suc h that f ( p i ) = q i It is natural to ask the question, do es suc h a dieomorphic mapping alw a ys exist? Our in tuition is it exist and there are innitely man y suc h dieomorphisms. This is stated as our rst theorem and the pro of follo ws. 42

PAGE 54

43 Theorem 5. A die omorphic tr ansformation that interp olates arbitr ary numb er of n p airs of landmarks always exists. Pr o of. W e sho w the existence b y construction. W e construct a simple, although most lik ely undesirable in most of the applications, dieomorphism. The in tuitiv e idea is to \dig canals" connecting the landmark pairs. W e rst c ho ose the rst pair of landmarks p 1 and q 1 Assume no other landmarks lie on the line connecting p 1 and q 1 Establish a co ordinate system suc h that p 1 and q 1 are on the x axis. Let the signed distance from p 1 to q 1 b e d Construct the transformation f 1 : n 1 n 2 suc h that f 1 ( x; y ) = ( x 0 ; y 0 ), x 0 = x + de v 2 (3.1) y 0 = y where v = tan ( 2 y ), for an y arbitrarily small It is easy to sho w that f 1 is a dieomorphism and it maps p 1 to q 1 and k eeps all other landmarks q 2 ,..., q n xed. This is v ery m uc h lik e the ro w of viscous ruid in a tub e. Similarly w e can construct dieomorphism f i that maps p i to q i and k eeps all other landmarks xed, for i = 1 ; 2 ; :::; n The comp osition of this series of dieomorphisms f = f n :::f 2 f 1 (3.2) is also a dieomorphism and ob viously f maps p i to q i for i = 1 ; 2 ; :::; n If some landmark q k lies on the line of p i and q i w e can nd suc h a direction suc h that w e dra w a line l k through q k and there are no other landmarks on the line. Then w e mak e a dieomorphism h transp orting q k to a nearb y p oin t q 0 k along the line without mo ving an y other landmarks, using the same canal of viscous ruid tec hnique. Then w e mak e dieomorphism f i as describ ed b efore. After that, w e mo v e landmark q 0 k bac k to the old p osition with the in v erse of h 1 So w e use F i = h 1 f i h in place of f i

PAGE 55

44 Figure 3{1: Existence of a dieomorphic mapping 3.2 Symmetric Matc hing due to Time Rev ersibilit y Asymmetry exist in man y image and shap e matc hing situations. Supp ose w e ha v e a p oin t set f p i 2 R 2 j i = 1 ; 2 ; :::n g and another p oin t set f q i 2 R 2 j i = 1 ; 2 ; :::n g w e nd the dieomorphic mapping f : R 2 R 2 whic h minimize the energy functional E ( f ) sub ject to the constrain ts f ( p i ) = q i No w w e dene a rev erse problem, namely to nd a mapping g : R 2 R 2 whic h minimize the same energy functional E ( f ) sub ject to the constrain ts f ( p i ) = q i as sho wn in Figure 3{2 In general, dep ending on the ob jectiv e function to minimize, g 6 = f 1 This is called asymmetry of the matc hing. In some cases w e do ha v e g = f 1 The matc hing is called symmetric then and this is a nice prop ert y to ha v e. The follo wing theorem states that for the exact dieomorphic landmark matc hing case, the matc hing is symmetric due to the time rev ersibilit y of the ro w.

PAGE 56

45 g f Figure 3{2: Asymmetry of the matc hing Theorem 6. If ( x k ; 1) = y k and ( x; t ) and v ( x; t ) minimize the ener gy E = Z 1 0 Z n jj Lv ( x; t ) jj 2 dxdt ; then the inverse mapping maps the landmarks b ackwar d 1 ( y k ; 1) = x k and 1 ( x; t ) and v ( x; t ) also minimize the ener gy E Pr o of. First, from the kno wn prop ert y of the dieomorphism group of suc h a dynamical system, ( x; t 1 + t 2 ) = ( ( x; t 1 ) ; t 2 ), it is easy to sho w that 1 ( x; t ) = ( x; t ). This is b ecause ( :; t ) ( :; t )( x ) = ( :; t ) ( :; t )( x ) = ( ( x; t ) ; t ) = ( x; t + ( t )) = ( x; 0) = x And ( x; t )and v ( x; t ) also satisfy the transp ort equation @ ( x; t ) @ t = v ( ( x; t ) ; t ). Supp ose ( x; t ) and v ( x; t ) minimize the energy E = R 1 0 R n jj Lv ( x; t ) jj 2 dxdt but 1 ( x; t ) = ( x; t ) and v ( x; t ) do not minimize the energy E = R 1 0 R n jj Lv ( x; t ) jj 2 dxdt Let the minimizer b e ( x; t ) and u ( x; t ) suc h that 8 k ( y k ) = x k and R 1 0 R n jj Lu ( x; t ) jj 2 dxdt < R 1 0 R n jj Lv ( x; t ) jj 2 dxdt Then, w e can construct 1 ( x; t ) = ( x; t ) suc h that 1 ( x; t ) and u ( x; t ) satisfy the transp ort equation and 1 ( x k ; 1) = y k Ho w ev er R 1 0 R n jj Lu ( x; t ) jj 2 dxdt < R 1 0 R n jj Lv ( x; t ) jj 2 dxdt con tradicts the assumption that v ( x; t ) is the minimizer of the energy E

PAGE 57

46 3.3 A Theoretical F ramew ork for Dieomorphic P oin t Matc hing There are dieren t w a ys to solv e for the unkno wn p oin t corresp ondences [ 56 14 17 16 ]. Essen tially within the framew ork of explicit p oin t corresp ondences|as opp osed to the distance function framew ork of implicit corresp ondence|w e ha v e a c hoice b et w een i) solving for an optimal p erm utation and ii) letting corresp onding \lab eled" p oin ts disco v er their optimal lo cations. W e opt for the latter in this w ork b ecause of its simplicit y The clustering in fact serv es t w o purp oses. First, it is the metho d to nd the unkno wn corresp ondence. W e initialize the t w o sets of cluster cen ters around the cen troids of their data p oin ts, resp ectiv ely The cluster cen ters are lab eled with iden tical lab els in the t w o sets denoting corresp ondence. The cluster cen ters ev olv e during the iterations of an incremen tal EM algorithm and they are link ed b y a dieomorphism and are forced to mo v e in lo c k-step with one another. Second, clustering is the mo deling of the real data sets, with noise and/or outliers b ecause with t w o shap es represen ted b y p oin t samples, w e cannot assume a p oin t-wise corresp ondence. The corresp ondence is only b et w een the t w o shap es and clustering is a useful w a y to mo del the shap es. W e use a Gaussian mixture mo del to describ e the clustering of the p oin t sets. F or more details on this approac h along with justications for the use of this mo del, please see [ 15 17 ]. The Gaussian mixture probabilit y densit y is p ( x j r ; T ) = 1 N N X k =1 1 (2 2 T ) d= 2 exp( 1 2 2 T jj x r k jj 2 ) (3.3) with x b eing a p oin t in R d r as the collectiv e notation of a set of cluster cen ters and 2 T as the v ariance of eac h Gaussian distribution. The reason for the notation of subscript T will b e discussed in Section 3.4 in the con text of annealing ( T will b e the temp erature. T and T are related b y T = 2 2 T .) Here w e just understand 2 T as the pr escrib e d v ariance in the Gaussian mixture mo del as opp osed to the actual measured v ariance 2 from data tting.

PAGE 58

47 The clustering pro cess is the estimation of the parameters r that leads to the maxim um log-lik eliho o d of the observ ed sample log p ( x j r ; T ) = N 1 X i =1 log N X k =1 exp( 1 2 2 jj x i r k jj 2 ) : (3.4) The solution can b e found b y applying the EM algorithm. As p oin ted out b y Hatha w a y [ 34 ], in the mixture mo del con text, the EM algorithm maximizing ( 3.4 ) can b e view ed as an alternativ e maximization of the follo wing ob jectiv e F ( M ; r ) = 1 2 2 T N 1 X i =1 N X k =1 M x ik jj x i r k jj 2 (3.5) N 1 X i =1 N X k =1 M x ik log M x ik : This is equiv alen t to minimizing E ( M ; r ) = F ( M ; r ) (3.6) = 1 2 2 T N 1 X i =1 N X k =1 M x ik jj x i r k jj 2 + N 1 X i =1 N X k =1 M x ik log M x ik with simplex constrain ts on M The clustering of the other p oin t set is iden tical. F or the join t clustering and dieomorphism estimation, w e put together the clustering energy of the t w o p oin t sets and the dieomorphic deformation energy induced in space giving us an

PAGE 59

48 ob jectiv e function E ( M x ; M y ; r ; s ; v ; ) (3.7) = 1 2 2 T N 1 X i =1 N X k =1 M x ik jj x i r k jj 2 + N 1 X i =1 N X k =1 M x ik log M x ik + 1 2 2 T N 2 X j =1 N X k =1 M y j k jj y j s k jj 2 + N 2 X j =1 N X k =1 M y j k log M y j k + 1 2 2 T N X k =1 jj s k ( r k ; 1) jj 2 + Z 1 0 Z n jj Lv ( x; t ) jj 2 dxdt: In the ab o v e ob jectiv e function, the cluster mem b ership matrices satisfy M x ik 2 [0 ; 1], 8 ik M y j k 2 [0 ; 1], 8 j k and P Nk =1 M 1 ik = 1, P Nk =1 M 2 j k = 1. The matrix en try M x ik is the mem b ership of data p oin t x i in cluster k whose cen ter is at lo cation r k The matrix en try M y j k is the mem b ership of data p oin t y j in cluster k whose cen ter is at p osition s k The dieomorphic deformation energy in n is induced b y the landmark displacemen ts from r to s where x 2 n and ( x; t ) is a one parameter dieomorphism: n n. Since the original p oin t sets dier in p oin t coun t and are unlab eled, w e cannot immediately use the dieomorphism ob jectiv e functions as in Joshi and Miller [ 39 ] or Camion and Y ounes [ 7 ] resp ectiv ely Instead, the t w o p oin t sets are clustered and the landmark dieomorphism ob jectiv e is used b et w een t w o sets of cluster cen ters r and s whose indices are alw a ys in corresp ondence. The dieomorphism ( x; t ) is generated b y the ro w v ( x; t ). ( x; t ) and v ( x; t ) together satisfy the transp ort equation @ ( x;t ) @ t = v ( ( x; t ) ; t ) and the initial condition 8 x ( x; 0) = x holds. This is in the inexact matc hing form and the displacemen t term P Nk =1 jj s k ( r k ; 1) jj 2 pla ys an imp ortan t role here as the bridge b et w een the t w o systems. This is also the reason wh y w e prefer the deformation energy in this form b ecause the coupling of the t w o sets of clusters app ear naturally through the inexact matc hing term and w e don't ha v e to in tro duce external coupling terms as

PAGE 60

49 in Guo et al. [ 33 ]. Another adv an tage of this approac h is that in this dynamic system describ ed b y the dieomorphic group ( x; t ), the landmarks trace a tra jectory exactly on the ro w lines dictated b y the eld v ( x; t ). Also, the feedbac k coupling is no longer needed as in the previous approac h b ecause with this deformation energy describ ed ab o v e, if ( x; t ) is the minimizer of this energy then 1 ( x; t ) is the bac kw ard mapping whic h also minimizes the same energy The T in ( 3.7 ) is a xed parameter. It is not a v ariable during the minimization. It is an attribute of the p oin t set and can b e a priori estimated for the p oin t shap es. The reason that w e ha v e co ecien t 1 2 2 T in fron t of the term P Nk =1 jj s k ( r k ; 1) jj 2 instead of another free parameter is that T is a natural unit of measuremen t for distan t discrepancies in clustering and it do es not mak e sense to mak e this co ecien t to o big or to o small. Since T is a constan t, w e m ultiply the ob jectiv e function b y a constan t 2 2 T and w e get the nal form of the ob jectiv e function as E ( M x ; M y ; r ; s ; v ; ) (3.8) = N 1 X i =1 N X k =1 M x ik jj x i r k jj 2 + 2 2 T N 1 X i =1 N X k =1 M x ik log M x ik + X N X k =1 M y j k jj y j s k jj 2 + 2 2 T N 2 X j =1 N X k =1 M y j k log M y j k + N X k =1 jj s k ( r k ; 1) jj 2 + 2 2 T Z 1 0 Z n jj Lv ( x; t ) jj 2 dxdt: 3.4 A Dieomorphic P oin t Matc hing Algorithm Our join t clustering and dieomorphism estimation algorithm has t w o comp onen ts: i) clustering and ii) dieomorphism estimation. F or the clustering part, w e use the deterministic annealing approac h. The clustering problem is a non-con v ex optimization problem. The traditional clustering

PAGE 61

50 tec hniques use descen t based algorithms and they tend to get trapp ed in a lo cal minim um. Rose et al. [ 59 ] prop osed an annealing approac h using analogies to statistical ph ysics. The clustering cost function is seen as the free energy of a Gibbs canonical distribution. The minimization of clustering cost function is seen as the sim ulation of a ph ysical annealing pro cess in whic h free energy is minimized. Let T b e the temp erature of the system and in the clustering system T = 2 2 T Let = 1 =T b e the recipro cal temp erature. Initially let = 0. W e ha v e a single minim um for the energy in ( 3.6 ), whic h is the global minim um and all the cluster cen ters are lo cated at the same p oin t, whic h is the cen ter of mass of all the data p oin ts and eac h data p oin t is uniformly asso ciated with all clusters. In the n umerical implemen tation, w e initialize all the cluster cen ters on a sphere of a v ery small radius and there are no data p oin t within the sphere. When the temp erature is lo w ered gradually at a certain critical v alue of temp erature, the clusters will split in to smaller clusters and a phase transition o ccurs. A t lo w er T the free energy ma y ha v e man y lo cal minima but the annealing pro cess is able to trac k the global minim um. F or the dieomorphism estimation, w e expand the ro w eld in term of the k ernel K of the L op erator v ( x; t ) = N X k =1 k ( t ) K ( x; k ( t )) (3.9) where k ( t ) is notational shorthand for ( r k ; t ) and w e also tak e in to consideration the ane part of the mapping (not written out in the ab o v e equation) when w e use the thin-plate k ernel with matrix en try K ij = r 2 ij log r ij for 2D and K ij = r ij for 3D, with r ij = k x i x j k After discretizing in time t the ob jectiv e in ( 3.7 ) is expressed as

PAGE 62

51 E ( M x ; M y ; r ; s ; ( t ) ; ( t )) (3.10) = N 1 X i =1 N X k =1 M x ik jj x i r k jj 2 + T N 1 X i =1 N X k =1 M x ik log M x ik + N 2 X j =1 N X k =1 M y j k jj y j s k jj 2 + T N 2 X j =1 N X k =1 M y j k log M y j k + N X k =1 k s k r k N X l =1 S X t =0 [ P ( t ) d l ( t ) + l ( t ) K ( k ( t ) ; l ( t ))] k 2 + T N X k =1 N X l =1 S X t =0 < k ( t ) ; l ( t ) > K ( k ( t ) ; l ( t )) where P ( t ) = 266666666664 1 11 ( t ) 21 ( t ) 31 ( t ) : : : : : : : : : : : : 1 1N ( t ) 2N ( t ) 3N ( t ) 377777777775 (3.11) and d is the ane parameter matrix. After w e p erform a QR decomp osition on P P ( t ) = ( Q 1 ( t ) : Q 2 ( t )) 0B@ R ( t ) 0 1CA : (3.12) W e iterativ ely solv e for k ( t ) and k ( t ) using an alternating algorithm. When k ( t ) is held xed, w e solv e for k ( t ). The solutions are d ( t ) = R 1 ( t ) [ Q 1 ( t ) ( t + 1) Q 1 ( t ) K ( ( t )) Q 2 ( t ) r ( t )] (3.13) ( t ) = Q 2 ( t ) r ( t ) (3.14)

PAGE 63

52 where K ( ( t )) denotes the thin-plate k ernel matrix ev aluated at ( t ) def = f ( r k ; t ) j k = 1 ; : : : ; N g and r ( t ) = ( Q T2 ( t ) K ( ( t )) Q 2 ( t ) + T ) 1 Q T2 ( t ) ( t + 1) : (3.15) When k ( t ) is held xed, w e use gradien t descen t to solv e for k ( t ): @ E @ k ( t ) = 2 N X l =1 < k ( t ) ; l ( t ) 2 W l > 5 1 K ( k ( t ) ; l ( t )) (3.16) where W l = s l r l P Nm =1 R 1 0 m ( t ) K ( m ( t ) ; l ( t )) dt The clustering of the t w o p oin t sets is handled b y a deterministic annealing EM algorithm whic h iterativ ely estimates the cluster mem b erships M x and M y and the cluster cen ters r and s The up date of the mem b erships is the v ery standard E-step of the EM algorithm [ 17 ] and is p erformed as sho wn b elo w. M x ik = exp( k x i r k k 2 ) P Nl =1 exp( k x i r l k 2 ) ; 8 ik (3.17) M y j k = exp( k y j s k k 2 ) P Nl =1 exp( k y j s l k 2 ) ; 8 j k : (3.18) The cluster cen ter up date is the M-step of the EM algorithm. This step is not the t ypical M-step. W e use a closed-form solution for the cluster cen ters whic h is an appro ximation. F rom the clustering standp oin t, w e assume that the c hange in the dieomorphism at eac h iteration is suciently smal l so that it c an b e ne gle cte d After making this appro ximation, w e get r k = P N 1 i =1 M x ik x k + s k P Nl =1 R 1 0 l ( t ) K ( l ( t ) ; k ( t )) dt 1 + P N 1 i =1 M x ik ; (3.19) s k = P N 2 j =1 M y j k y j + ( r k ; 1) 1 + P N 2 j =1 M y j k ; 8 k : (3.20) The o v erall algorithm is describ ed b elo w.

PAGE 64

53 Initialization: Initial temp erature T = 0 : 5(max i k x i x c k 2 + max j k y j y c k 2 ) where x c and y c are the cen troids of X and Y resp ectiv ely Begin A: While T > T nal Step 1 : Clustering Up date mem b erships according to ( 3.17 ), ( 3.18 ). Up date cluster cen ters according to ( 3.19 ), ( 3.20 ). Step 2: Dieomorphism Up date ( ; v ) b y minimizing E di ( ; v ) = C X k =1 jj s k ( r k ; 1) jj 2 + T Z 1 0 Z n jj Lv ( x; t ) jj 2 dxdt according to ( 3.13 )( 3.14 ) and ( 3.16 ). Step 3: Annealing. T r T where r < 1. End 3.5 Applications to 2D Corpus Callosum Shap es W e applied the algorithm to nine sets of 2D corpus callosum slices. The feature p oin ts w ere extracted with the help of a neuroanatomical exp ert. Figure 3{3 sho ws the nine corpus callosum 2D images, lab eled CC1 through CC9. In our exp erimen ts, w e rst did the sim ultaneous clustering and matc hing with the corpus callosum p oin t sets CC5 and CC9. The clustering of the t w o p oin t sets is sho wn in Figure 3{4 There are 68 cluster cen ters. The circles represen t the cen ters and the dots are the data p oin ts. The t w o cluster cen ters induce the dieomorphic mapping of the 2D space. The w arping of the 2D grid under this dieomorphism is sho wn in Figure 3{5 Using this dieomorphism, w e calculated the after-image of original data p oin ts and compared them with the target data p oin ts. Due to the

PAGE 65

54 large n um b er of cluster cen ters, the cluster cen ters nearly coincide with the original data p oin ts and the w arping of the original data p oin ts is not sho wn in the gure. The corresp ondences (at the cluster lev el) are sho wn in Figure 3{6 The algorithm allo ws us to sim ultaneously obtain the dieomorphism and the corresp ondence. Using our form ulation, w e are able to calculate the geo desic distances b et w een the t w o sets of cluster cen ters. This is done on the shap e manifold. Eac h p oin t q on the shap e manifold M represen ts a set of N cluster cen ters x 1 x 2 ... x N 2 R 2 and has a co ordinate q = ( x 11 ; x 21 ; x 12 ; x 22 ; :::; x 1N ; x 2N ) where x i = ( x 1i ; x 2i ), i = 1 ; 2 ; :::; N Let q ( t ) b e the geo desic path connecting t w o p oin ts q 1 and q 2 on the manifold. Using the norm dened for the tangen t v ector in Camion and Y ounes [ 7 ], the geo desic distance b et w een q 1 and q 2 is D geo desic ( q 1 ; q 2 ) = Z 1 0 q q T Q 2 ( Q T2 K ( q ) Q 2 + ) 1 Q T2 q dt: (3.21) where K ( q ) is the k ernel of the L op erator ev aluated at q ( t ) and as men tioned previously the thin-plate spline k ernel is used. Q 2 comes from the QR decomp osition of P P = ( Q 1 : Q 2 ) 0B@ R 0 1CA (3.22) and P = 0BBBBBBBBBB@ 1 x 11 x 21 : : : : : : : : : 1 x 1N x 2N 1CCCCCCCCCCA (3.23) W e also exp erimen ted with dieren t n um b er of cluster cen ters. T able 3{1 sho ws a mo died Hausdor distance as rst in tro duced [ 20 ] b et w een the image

PAGE 66

55 set of p oin ts CC5 after dieomorphism and the target set of p oin ts CC9 when the n um b er of clusters v ary The reason for using the mo died Hausdor distance instead of the Hausdor distance is that the latter is to o sensitiv e to outliers. The denition of the mo died Hausdor distance is H mo d ( A; B ) = max ( h mo d ( A; B ) ; h mo d ( B ; A )) ; (3.24) where A and B are nite p oin t sets and h mo d ( A; B ) = 1 j A j X a 2 A min b 2 B k a b k (3.25) is the a v erage of the minim um distances instead of the maxim um of the minim um distances. It is easy to see that when the n um b er of clusters increases, the matc hing impro v es as the mo died Hausdor distance decreases. In the third column in T able 3{2 w e list the geo desic distances b et w een the t w o sets of cluster cen ters after pair-wise w arping and clustering all the pair of corpus callosum p oin t sets. Using the cluster cen ters as landmarks, a dieomorphic mapping of the space is induced. With this induced dieomorphism, w e mapp ed the original data sets and compared the image of the original p oin t set under the dieomorphism and the target p oin t set using the mo died Hausdor distance. The mo died Hausdor distances b et w een the pairs are listed in the fourth column in T able 3{2 Finally from the original nine corpus callosum p oin t sets, w e w arp ed the rst eigh t p oin t sets on to the nin th set and Figure 3{7 displa ys the o v erla y of all p oin t sets after dieomorphic w arping. 3.6 Applications to 3D Shap es W e applied our form ulation and algorithm to the 3D p oin t data of hipp o campal shap es. W e rst applied the algorithm to syn thetic data, where w e ha v e the

PAGE 67

56 0.4 0.6 0.8 1 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.2 0 0.2 Figure 3{3: P oin t sets of nine corpus callosum images. 0.4 0.6 0.8 1 1.2 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5 0.4 0.6 0.8 1 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5 Figure 3{4: Clustering of the t w o p oin t sets. 0.2 0.4 0.6 0.8 1 1.2 0.2 0.1 0 0.1 0.2 0.3 0.4 0.2 0.4 0.6 0.8 1 1.2 0.2 0.1 0 0.1 0.2 0.3 0.4 Figure 3{5: Dieomorphic mapping of the space.

PAGE 68

57 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.1 0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Figure 3{6: Matc hing b et w een the t w o p oin t sets. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.1 0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 Figure 3{7: Ov erla y of the after-images of eigh t p oin t sets with the nin th set. T able 3{1: Mo died Hausdor distance of the matc hing p oin t sets. Num b er of Clusters Mo died Hausdor Distance 10 0.0082 20 0.0082 30 0.0057 40 0.0050 50 0.0043 60 0.0035 68 0.0027

PAGE 69

58 T able 3{2: Geo desic distances b et w een t w o sets of cluster cen ters and mo died Hausdor distances of matc hing p oin ts. F rom T o Geo desic M.H. CC1 CC2 0.0264 0.0055 CC1 CC3 0.0132 0.0014 CC1 CC4 0.0289 0.0048 CC1 CC5 0.0269 0.0056 CC1 CC6 0.0250 0.0097 CC1 CC7 0.0323 0.0054 CC1 CC8 0.0256 0.0043 CC1 CC9 0.0241 0.0041 CC2 CC3 0.0277 0.0059 CC2 CC4 0.0342 0.0063 CC2 CC5 0.0308 0.0057 CC2 CC6 0.0211 0.0100 CC2 CC7 0.0215 0.0040 CC2 CC8 0.0271 0.0044 CC2 CC9 0.0258 0.0093 CC3 CC4 0.0443 0.0059 CC3 CC5 0.0294 0.0047 CC3 CC6 0.0181 0.0032 CC3 CC7 0.0256 0.0060 CC3 CC8 0.0153 0.0018 CC3 CC9 0.0305 0.0044 CC4 CC5 0.0231 0.0046 CC4 CC6 0.0304 0.0056 CC4 CC7 0.0324 0.0056 CC4 CC8 0.0311 0.0054 CC4 CC9 0.0434 0.0090 CC5 CC6 0.0266 0.0056 CC5 CC7 0.0325 0.0053 CC5 CC8 0.0225 0.0037 CC5 CC9 0.0305 0.0069 CC6 CC7 0.0244 0.0050 CC6 CC8 0.0186 0.0026 CC6 CC9 0.0274 0.0056 CC7 CC8 0.0212 0.0044 CC7 CC9 0.0241 0.0102 CC8 CC9 0.0196 0.0050

PAGE 70

59 200 220 240 260 280 300 320 340 360 120 140 160 180 200 220 240 Figure 3{8: Tw o p oin t sets of hipp o campal shap es. The set with crosses is the original set and the set with dots is the one after GRBF w arping. kno wledge of ground truth and this serv es as the v alidation of the algorithm. W e then exp erimen ted with real data and ev aluated the results using v arious measures. 3.6.1 Exp erimen ts on Syn thetic Data W e selected one hipp o campal p oin t set and w arp ed it with a kno wn dieomorphism using the Gaussian Radial Basis F unction (GRBF) k ernel. W e c ho ose = 60 for the GRBF b ecause with this large v alue of ,w e are able to generate a more global w arping. Figure 3{8 sho ws the t w o p oin t sets of hipp o campal shap es. The set with crosses is the original set and the set with dots is the one after GRBF w arping. First, w e ha v e no noise added. W e used the TPS k ernel to reco v er the dieomorphism via join t clustering using our algorithm. The reason w e use dieren t k ernels for w arping and reco v ering is the ob jectiv eness. It is trivial to reco v er the

PAGE 71

60 T able 3{3: Matc hing errors on syn thetic data with dieren t noise lev els Noise lev el n No. clusters 100 200 300 400 500 0 0.21 0.17 0.16 0.19 0.20 0.1 0.30 0.13 0.28 0.09 0.26 0.08 0.29 0.09 0.31 0.11 0.2 0.41 0.16 0.39 0.12 0.35 0.11 0.37 0.13 0.39 0.14 0.3 0.44 0.17 0.41 0.13 0.39 0.15 0.40 0.16 0.42 0.19 0.4 0.61 0.23 0.54 0.19 0.52 0.18 0.55 0.20 0.59 0.21 0.5 0.68 0.24 0.62 0.25 0.59 0.24 0.63 0.22 0.65 0.28 0.6 0.82 0.38 0.75 0.35 0.72 0.33 0.76 0.37 0.80 0.36 0.7 0.96 0.49 0.90 0.42 0.86 0.42 0.90 0.44 0.94 0.46 0.8 1.21 0.54 1.13 0.51 0.92 0.48 1.09 0.49 1.18 0.51 0.9 1.63 0.72 1.48 0.66 1.45 0.62 1.49 0.61 1.52 0.68 1.0 1.82 0.78 1.70 0.71 1.64 0.67 1.69 0.73 1.77 0.75 deformation that is w arp ed with the same k ernel. Since the reference data are syn thesized, w e kno w the ground truth and w e are able to compare our result with the ground truth. After un w arping the p oin t set with our reco v ered dieomorphism, w e nd the squared distances b et w een the corresp onding data p oin ts, and nd the a v erage and then tak e the square ro ot. This is the standard error for our reco v ered dieomorphism. W e ha v e t w o free parameters, and T f inal T f inal is determined b y the limiting v alue of T whic h is in turn determined b y the n um b er of clusters. W e c ho ose a v alue suc h that the whole optimization pro cess is stable in the temp erature range from initial T to T f inal W e exp erimen ted with dieren t n um b ers of clusters and listed the corresp onding standard errors in the rst ro w of T able 3{3 It is easy to see that the standard error go es do wn as the n um b er of clusters go es up from 100 to 300 and go es up again when the n um b er of clusters increases further. This is b ecause when w e ha v e to o few clusters, the p oin ts are not w ell represen ted b y the cluster cen ters. On the other hand, if w e ha v e to o man y clusters, the v ariance b et w een the t w o shap es is to o big and the deformation increases dramatically F or the standard error, there is an optimal n um b er of clusters and in this case w e nd it to b e 300. W e need to estimate the macroscopic and microscopic dimensions of the shap e in order to see ho w big the standard error is. W e calculated the co v ariance matrix of the original data set. W e nd their

PAGE 72

61 T able 3{4: Limiting v alue of determined b y the n um b er of clusters Num b er of clusters 100 200 300 400 500 Limit 3.3 1.9 1.2 0.9 0.6 eigen v alues to b e [48.1, 11.8, 4.1]. This giv es us an estimate of the macroscopic dimensions to b e ab out 100, 24 and 8, namely t wice the eigen v alues. W e then nd out the a v erage distance b et w een the nearest neigh b ors to b e 2.65. This giv e us the microscopic dimension of the shap e. As w e can see from the table, our matc hing is v ery accurate. Next w e add noise to the w arp ed data and test the robustness of our algorithm to noise. After GRBF w arping, w e add Gaussian noise to the w arp ed data with dieren t v ariances W e exp erimen ted with ten trials for eac h noise lev el from 0.1 to 1.0 and for eac h cluster lev el from 100 to 500. The standard errors and the deviation are sho wn in T able 3{3 W e can see the standard error increase with the increasing noise lev el but it appro ximately sta ys in the range of the noise. Stronger noise do es not increase the matc hing error dramatically and this sho ws the algorithm is robust against noise. This is easier to see when plotted in Figure 3{9 with error bars. W e split the v e lev els of clusters in to t w o plots b ecause it lo oks messy if they w ere put together in a single plot Figure 3{9 (a) has the errors for 100, 200 and 300 clusters and Figure 3{9 (b) has the errors for 300, 400 and 500 clusters. W e can see that at the 300 cluster lev el, w e obtain the b est matc hing. 3.6.2 Exp erimen ts on Real Data W e applied the algorithm on dieren t real hipp o campal data sets. Figure 3{10 sho ws t w o hipp o campal shap es. Figure 3{11 sho ws the annealing pro cess. The x axis is the iteration step. The dashed line is the scaled temp erature p T = 2 or T The solid line is the actual v ariance W e can see when the temp erature go es do wn, it driv es the do wn. W e observ e a phase transition at temp erature

PAGE 73

62 T = 3 : 37 10 3 W e observ e that there is a lo w er limit for When the temp erature gets v ery lo w, the b ecomes a constan t, whic h is 1.2, and no matter ho w m uc h lo w er the temp erature gets, the sta ys constan t. This constan t is determined b y the n um b er of clusters. In Figure 3{12 a through Figure 3{12 e, w e sho w ho w this limit c hanges with the n um b er of clusters. When the n um b er of clusters equals or exceeds the n um b er of data p oin ts, the limit approac hes zero. T able 3{4 displa ys the limits of c hanging with the n um b er of clusters. Because of noise and sampling error, w e should not allo w this limit to go to zero. Again w e observ e when w e ha v e 300 clusters, w e ha v e a reasonable = 1 : 2 as w e recall the a v erage distance b et w een the nearest neigh b ors is ab out 2.65. Figure 3{13 sho ws the clustering of the t w o shap es. W e then did the matc hing for all the pairs out of ten hipp o campal shap es. T able 3{5 and T able 3{6 sho ws three measures for the matc hing results with dieren t clusters: Jensen-Shannon div ergence, Hausdor distance and mo died Hausdor distance. The Jensen-Shannon div ergence (a sp ecial case with = 1 = 2) is dened as [ 25 ] D = Z n ( p ( x ) log 2 p ( x ) p ( x ) + q ( x ) + q ( x ) log 2 q ( x ) p ( x ) + ( x ) ) d x ; (3.26) where x is the random v ariable while p ( x ) and q ( x ) are the t w o probabilit y densities. Notice this measure is highly non-linear. When p ( x ) and q ( x ) are completely indep enden t, namely in our matc hing case, when the t w o shap es are completely dieren t, D has a maxim um of 2 log 2 = 1 : 39. In practice, w e use the follo wing tec hnique to compute D W e observ e in ( 3.26 ), the in tegral can b e expressed as the exp ectation v alues of some functions under t w o dieren t probabilit y distributions: D = < log 2 p ( x ) p ( x ) + q ( x ) > p + < log 2 q ( x ) p ( x ) + ( x ) > q ; (3.27)

PAGE 74

63 where < log 2 p ( x ) p ( x )+ q ( x ) > p is the exp ectation v alue of function log 2 p ( x ) p ( x )+ q ( x ) under probabilit y distribution p ( x ) and < log 2 q ( x ) p ( x )+( x ) ) > q is the exp ectation v alue of function log 2 q ( x ) p ( x )+( x ) ) under probabilit y distribution q ( x ). In our Gaussian mixture mo del, w e see the data p oin ts as samples from a Gaussian mixture probabilit y distribution with kno wn cluster cen ters. Here p ( x ) = 1 N 1 (2 2 ) 3 = 2 N X k =1 exp( jj x r k jj 2 2 2 ) ; (3.28) and q ( x ) = 1 N 1 (2 2 ) 3 = 2 N X k =1 exp( jj x s k jj 2 2 2 ) ; (3.29) where x is the random v ariable, namely the space lo cation and f r k g is the rst set of cluster cen ters and f s k g is the second set of cluster cen ters. W e use the a v erage of nite samples as an appro ximation of the exp ectation v alues and w e ha v e D = 1 N 1 N 1 X i =1 log 2 p ( x i ) p ( x i ) + q ( x i ) + 1 N 2 N 2 X j =1 log 2 q ( y j ) p ( y j ) + ( y j ) ; (3.30) where f x i g is the rst set of data p oin ts; N 1 is the n um b er of p oin ts in the rst set; f y j g is the second set of data p oin ts; and N 2 is the n um b er of p oin ts in the second set. W e ha v e seen that Jensen-Shannon div ergence is v ery useful in estimating the v alidit y of the registration of t w o p oin t shap es without kno wing the ground truth. The Hausdor distance is dened as H ( A; B ) = max( h ( A; B ) ; h ( B ; A )) ; (3.31) where A and B are nite p oin t sets and h ( A; B ) = max a 2 A min b 2 B k a b k : (3.32) The Hausdor distance measures the w orst case dierence b et w een the t w o p oin t sets. F rom T able 3{5 and T able 3{6 w e can see that when w e ha v e 300

PAGE 75

64 clusters, w e ha v e the minim um Jensen-Shannon div ergence and the Hausdor distance. Ho w ev er, the Hausdor distance is to o sensitiv e to outliers. W e also calculated the mo died Hausdor distance as rst in tro duced in Dubuisson and Jain [ 20 ]. The denition of the mo died Hausdor distance w as giv en b efore in ( 3.24 ) and ( 3.25 ). It is the a v erage of the minim um distances instead of the maxim um of the minim um distances. It is easy to see that when the n um b er of clusters increases, the mo died Hausdor distance decreases.

PAGE 76

65 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.5 1 1.5 2 2.5 3 N=100 N=200 N=300 (a) 0.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.5 1 1.5 2 2.5 3 N=500 N=400 N=300 (b) Figure 3{9: Matc hing errors on syn thetic data for dieren t n um b er of clusters

PAGE 77

66 220 240 260 280 300 320 340 360 380 120 140 160 180 200 220 240 0 10 20 Figure 3{10: Tw o hipp o campal shap es

PAGE 78

67 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 iteration steps s T s x100 Figure 3{11: Deterministic annealing in the clustering pro cess: the dashed line is the scaled temp erature p T = 2 or T The solid line is the actual v ariance When the temp erature go es do wn, it driv es the do wn. There is a phase transition at temp erature T = 3 : 37 10 3 and there exists a lo w er limit 1.2 for

PAGE 79

68 40 50 60 70 80 90 100 0 1 2 3 4 5 6 7 8 9 10 11 12 iteration steps s s T (a) 100 clusters 40 50 60 70 80 90 100 0 1 2 3 4 5 6 7 8 9 10 11 12 iteration steps s s T (b) 200 clusters 40 50 60 70 80 90 100 0 1 2 3 4 5 6 7 8 9 10 11 12 iteration steps s s T (c) 300 clusters 40 50 60 70 80 90 100 0 1 2 3 4 5 6 7 8 9 10 11 12 iteration steps s s T (d) 400 clusters 40 50 60 70 80 90 100 0 1 2 3 4 5 6 7 8 9 10 11 12 iteration steps s s T (e) 500 clusters Figure 3{12: Limiting v alue of determined b y the n um b er of clusters

PAGE 80

69 (a) Clustering of the template hipp o campal shap e (b) Clustering of the reference hipp o campal shap e Figure 3{13: Clustering of the t w o hipp o campal shap es

PAGE 81

70 T able 3{5: Jensen-Shannon div ergence for v arious pairs of shap es Jensen-Shannon div. T rial no. n No. clusters 100 200 300 400 500 1 0.87 0.31 0.03 0.13 0.21 2 0.93 0.62 0.47 0.05 0.24 3 0.76 0.27 0.04 0.16 0.32 4 0.98 0.52 0.34 0.09 0.45 5 0.69 0.41 0.14 0.18 0.36 6 0.57 0.23 0.43 0.78 0.97 7 0.66 0.21 0.05 0.14 0.30 8 0.99 0.70 0.25 0.19 0.63 9 0.85 0.42 0.11 0.68 0.74 10 0.97 0.62 0.10 0.18 0.55 11 0.70 0.33 0.06 0.13 0.26 12 1.02 0.64 0.08 0.44 0.71 13 0.89 0.54 0.20 0.31 0.65 14 0.57 0.09 0.15 0.66 0.80 15 0.88 0.30 0.05 0.29 0.36 16 0.90 0.75 0.12 0.17 0.44 17 0.61 0.16 0.28 0.53 0.72 18 0.91 0.37 0.18 0.40 0.88 19 1.12 0.80 0.47 0.09 0.28 20 0.96 0.54 0.33 0.60 0.74 21 0.65 0.23 0.51 0.78 1.04 22 0.93 0.46 0.22 0.51 0.68 23 0.92 0.60 0.28 0.15 0.34 24 0.80 0.26 0.57 0.69 0.86 25 1.10 0.62 0.44 0.78 0.97 26 0.90 0.39 0.05 0.21 0.47 27 0.58 0.07 0.20 0.56 0.77 28 0.93 0.51 0.09 0.40 0.63 29 0.99 0.26 0.18 0.37 0.70 30 0.60 0.06 0.17 0.54 0.57 31 0.83 0.19 0.08 0.37 0.76 32 1.22 0.42 0.57 0.70 0.95 33 0.80 0.59 0.30 0.86 0.92 34 0.89 0.76 0.35 0.28 0.67 35 1.05 0.42 0.37 0.81 1.13 36 0.92 0.25 0.31 0.60 0.85 37 0.79 0.35 0.08 0.24 0.40 38 0.90 0.42 0.16 0.35 0.68 39 0.86 0.27 0.38 0.50 0.71 40 0.55 0.04 0.19 0.36 0.67 41 1.02 0.30 0.47 0.81 0.98 42 0.43 0.07 0.22 0.56 0.86 43 0.78 0.56 0.18 0.39 0.61 44 0.61 0.09 0.25 0.70 0.82 45 0.44 0.15 0.26 0.53 0.93

PAGE 82

71 T able 3{6: Hausdor and mo died Hausdor distance for v arious pairs of shap es Hausdor distance mo died Hausdor T rial no. n No. clusters 100 200 300 400 500 100 200 300 400 500 1 7.1 7.4 5.7 6.2 7.3 2.8 2.0 1.4 1.2 1.1 2 9.3 8.9 7.2 8.3 8.7 3.5 3.1 2.8 2.4 2.3 3 7.2 6.1 4.9 5.6 6.4 2.0 1.7 1.4 1.3 1.2 4 8.4 7.8 7.2 5.2 6.5 2.7 2.4 2.3 1.7 1.4 5 9.6 9.7 8.0 8.4 8.9 3.9 3.6 3.1 2.8 2.7 6 9.2 6.3 7.1 7.8 8.6 3.1 2.8 2.5 2.2 2.1 7 6.9 5.8 4.4 6.0 7.3 2.4 2.2 2.1 1.7 1.5 8 8.9 8.5 7.0 6.4 8.2 3.0 2.6 2.4 2.2 1.9 9 9.3 8.0 5.9 7.6 9.1 2.9 2.7 2.3 2.1 1.6 10 7.8 7.3 4.7 6.7 8.1 3.2 2.8 2.3 1.8 1.4 11 8.7 7.7 5.8 7.4 9.0 2.5 2.1 1.6 1.4 1.2 12 9.1 8.3 6.1 7.4 8.6 3.3 3.0 2.5 2.2 2.0 13 9.3 8.4 6.5 7.0 8.7 3.6 3.4 3.1 2.4 2.2 14 7.4 5.1 5.5 6.8 8.3 3.0 2.7 2.3 2.0 1.8 15 8.8 7.1 4.9 6.3 7.8 2.6 2.0 1.5 1.3 1.2 16 9.4 9.0 6.1 7.3 8.5 3.2 3.0 2.4 2.1 1.9 17 8.6 6.8 7.9 9.0 9.9 3.4 3.5 3.1 2.7 2.4 18 9.5 8.2 6.5 7.4 8.0 3.7 3.2 2.9 2.6 2.1 19 9.2 7.8 7.2 5.1 6.4 2.8 2.6 2.3 2.2 2.0 20 9.6 8.0 7.3 8.7 9.3 3.9 3.5 3.3 3.1 2.7 21 8.4 6.1 6.9 7.8 9.5 3.3 3.1 2.8 2.4 2.3 22 9.7 8.5 7.0 8.1 9.0 2.9 2.7 2.6 2.3 2.1 23 9.6 8.2 7.3 6.5 7.7 2.4 2.1 1.7 1.6 1.5 24 7.8 6.6 7.2 8.9 9.6 3.1 2.7 2.5 2.2 1.9 25 9.8 7.9 7.6 8.8 9.2 3.8 3.4 3.0 2.8 2.7 26 9.0 7.3 5.8 7.0 8.7 2.9 2.4 2.0 1.4 1.2 27 7.8 6.0 6.5 7.2 8.3 3.2 2.9 2.5 2.1 1.9 28 9.5 8.1 6.1 7.4 8.8 3.0 2.8 2.5 2.1 1.8 29 9.7 8.3 6.7 7.0 8.5 3.4 2.7 2.2 1.9 1.7 30 7.1 5.6 6.2 7.3 7.8 2.5 2.2 1.8 1.3 1.2 31 9.3 6.7 5.8 7.5 8.4 2.7 2.4 2.2 1.6 1.4 32 9.9 7.2 7.8 8.5 9.2 3.5 3.0 2.8 2.5 2.3 33 8.9 8.0 6.3 7.1 8.7 3.3 3.1 2.6 2.4 2.3 34 8.9 8.5 6.7 6.4 7.0 3.0 2.8 2.5 2.2 2.1 35 9.6 8.4 7.0 8.1 9.9 3.7 3.4 3.0 2.3 2.0 36 8.7 5.9 6.2 7.5 8.4 3.2 2.9 2.5 2.3 2.2 37 7.7 6.3 5.4 6.1 7.6 2.4 2.2 1.9 1.6 1.5 38 9.5 7.1 5.7 6.6 7.9 2.8 2.4 2.3 2.0 1.8 39 9.2 7.3 8.2 8.4 9.0 3.2 3.0 2.7 2.5 2.4 40 6.5 5.2 5.8 6.7 7.3 2.0 1.6 1.3 1.1 0.9 41 9.4 7.5 7.9 8.6 9.2 3.8 3.2 2.5 2.4 2.3 42 7.2 5.8 6.7 7.3 7.9 2.3 1.9 1.6 1.5 1.4 43 8.2 7.4 6.0 6.7 7.8 2.9 2.4 2.2 2.0 1.8 44 7.1 5.9 6.6 7.5 8.0 2.2 1.9 1.7 1.5 1.4 45 7.0 6.5 7.2 7.7 9.1 2.5 2.1 1.9 1.7 1.6

PAGE 83

CHAPTER 4 TOPOLOGICAL CLUSTERING AND MA TCHING In this c hapter w e extend our dieomorphic p oin t matc hing theory and algorithm to include the kno wn information of the top ology of the underlying shap es. In Section 4.1 w e pro vide a brief in tro duction to the basic concepts of top ological spaces. In Section 4.2 w e review the Kohonen Self-Organizing F eature Map (SOFM) whic h w as rst in tro duced b y Kohonen in the con text of neural net w orks in 1980s. Our top ological clustering and matc hing is related to SOFM b ecause the essence of SOFM is top ology preserving. Ho w ev er, our top ological clustering and matc hing is dieren t from SOFM in man y w a ys and as part of Section 4.3 w e discuss these dierences. In Section 4.3 w e discuss the motiv ation and metho ds of top ological clustering and matc hing. W e do this with graph top ology assigned to the set of cluster cen ters. The graph top ology can b e prescrib ed if w e ha v e prior kno wledge of the shap e top ology or it can b e arbitrary in the case w e don't kno w the shap e top ology in adv ance. In Section 4.5 w e presen t results for clustering and matc hing with prescrib ed top ology with example of c hain top ology ring top ology and gen us zero closed surface top ology or S 2 top ology Section 4.6 describ es ho w w e can appro ximate the top ology if w e do not ha v e the prior kno wledge ab out the top ology of the shap e in adv ance. 4.1 F undamen tals of T op ological Spaces The heart of top ology is the concept of \nearness," describ ed b y \neigh b orho o d". W e rst in tro duce the concept of top ological space. Readers can confer the Encyclop edic Dictionary of Mathematics compiled b y Mathematical So ciet y of Japan and translated b y Massac h usetts Institute of T ec hnology [ 47 ]. W e also cite from a b o ok b y Bourbaki [ 6 ] and a b o ok b y Rosenfeld [ 61 ]. 72

PAGE 84

73 F elix Hausdor in his F oundations of Set The ory (Grundz uge der Mengenlehre. Leipzig, 1914) [ 35 36 ] dened his concept of a top ological space based on the four axioms. Let X b e a set. A neigh b orho o d system for X is a function U that assigns to eac h p oin t x of X a family U ( x ) of subsets of X sub ject to the follo wing axioms ( U ): ( U 1) x 2 U for eac h U in U ( x ). ( U 2) If U 1 U 2 2 U ( x ), then U 1 \ U 2 2 U ( x ). ( U 3) If U 2 U ( x ) and U V then V 2 U ( x ). ( U 4) F or eac h U in U ( x ), there is a mem b er W of U ( x ) suc h that U 2 U ( y ) for eac h y in W U ( x ) is in terpreted as the family of all neigh b orho o ds of p oin t x An elemen t U 2 U ( x ) is called a neigh b orho o d of p oin t x The in tuitiv e translation of the ab o v e axioms is as follo ws. ( U 1) x is in eac h neigh b orho o d of x ( U 2) The in tersection of t w o neigh b orho o ds of x is a neigh b orho o d of x ( U 3) If a set V con tains a neigh b orho o d of x then V is itself a neigh b orho o d of x ( U 4) F or eac h neigh b orho o d U of x there is another neigh b orho o d W of x suc h that U is the neigh b orho o d of eac h p oin t y in W P a v el Sergeevi c Aleksandro v [Alexandro] prop osed in the pap er On the foundation of n-dimensional top olo gy (Zur Begr undugn der n -dimensionalen T op ologie. Leipzig, 1925) [ 1 ]: A system of op en sets for a set X is a family O of subsets of X satisfying the follo wing axioms ( O ): ( O 1) X ; 2 O ( O 2) If O 1 O 2 2 O then O 1 \ O 2 2 O

PAGE 85

74 ( O 3) If O 2 O ( 2 ), then S 2 O 2 O The elemen ts in O are called op en sets. An easy in tuitiv e in terpretation of this set of axioms is ( O 1) The empt y set is an op en set. The en tire space X is an op en set. ( O 2) The in tersection of t w o op en sets is an op en set. ( O 3) The union of arbitrarily man y op en sets is an op en set. Using the set complemen t and DeMorgan's la w w e get a system of closed sets. A system of closed sets for a space X is a family F of subsets of X satisfying the follo wing axioms ( F ): ( F 1) X ; 2 F ( F 2) If F 1 F 2 2 F then F 1 [ F 2 2 F ( F 3) If F 2 F ( 2 ), then T 2 F 2 F The elemen ts in F are called close d sets. An easy in tuitiv e in terpretation of this set of axioms is ( F 1) The en tire space X is a closed set. The empt y set is a closed set. ( F 2) The union of t w o closed sets is a closed set. ( F 3) The in tersection of arbitrarily man y closed sets is a closed set. Kazimierz Kurato wski in the pap er The op er ation A of analysis situs (L'op eration A de l'analysis situs. W arsa w, 1922) [ 44 ] prop osed: A closure op erator for a space X is a function that assigns to eac h subset A of X a subset A a of X satisfying the follo wing axioms( C ): ( C 1) ; a = ; ( C 2) ( A [ B ) a = A a [ B a ( C 3) A A a ( C 4) A a = A aa The in tuition of the closure of a set A is the union of A and its b oundary This set of axioms explained in w ords is as follo ws.

PAGE 86

75 ( C 1) The closure of the empt y set is the empt y set. ( C 2) The closure of ( A [ B ) is the union of closure of A and the closure of B ( C 3) The closure of A con tains A as a subset. ( C 4) The closure of the closure of A is the same as the closure of A This is sa ying that the closure op erator is idemp oten t. Related to the closure op erator is an in terior op erator. An in terior op erator for a space X is a function that assigns to eac h subset A of X a subset A i of X satisfying the follo wing axioms ( I ): ( I 1) X i = X ( I 2) ( A \ B ) i = A i \ B i ( I 3) A i A ( I 4) A ii = A i The in tuition of the in terior of a set A is A min us its b oundary This set of axioms explained in w ords is as follo ws. ( I 1) The in terior of the en tire space X is itself. ( I 2) The in terior of ( A \ B ) is the in tersection of the in terior of A and the in terior of B ( I 3) The in terior of A is a subset of A ( I 4) The in terior of the in terior of A is the same as the in terior of A This is sa ying that the in terior op erator is idemp oten t. All these systems equiv alen tly dene the top ological space. W e can in terpret one easily in the language of another. W e can dene op en set using neigh b orho o d: A set A X is an op en set, if 8 x 2 A ther e exists a neighb orho o d U of x such that U A then A is an op en set. Dene closed set using op en set: A set A X is a close d set, if X A is an op en set.

PAGE 87

76 Dene closure of A using closed set: If A X the closur e of A is the set A a T B wher e B is a close d set and A B In other wor ds, the closur e of A is the smal lest close d set that c ontains A Dene in terior using closure: A i X ( X A ) a In other wor d, for set A X rst nd the c omplement of A B = X A Then nd B a the closur e of B The the interior of A is the c omplement of B a Dene neigh b orho o d using in terior: A set U X is the neighb orho o d of p oint x 2 X if x 2 U i namely x is in the interior of U Figure 4{1 sho ws some examples of dieren t top ological spaces. Another imp ortance concept in top ological spaces is the separation axioms, dictating the exten t the p oin ts are separated from eac h other. Of our in terest is one particular space, called Hausdor space. Denition 4. A top olo gic al sp ac e is c al le d Hausdor sp ac e if any two distinct p oints have disjoint neighb orho o ds. 4.2 Kohonen Self-Organizing F eature Map (SOFM) Kohonen dev elop ed the Self-Organizing F eature Map algorithm. He describ ed the SOFM in the con text of neural net w orks with an aim to understand the the brain cortex mapping of sensory organs, lik e retina [ 41 40 42 ]. In the Kohonen net w ork, there are t w o la y ers of neurons, with the rst la y er the input and the second la y er the output. The output neurons are arranged in a rectangular grid. Eac h input neuron is connected with eac h output neuron. Eac h output neuron is asso ciated with d -w eigh ts, with d b eing the n um b er of the input neurons. Kohonen describ es a pro cedure of the initialization and up date of the w eigh ts. The imp ortan t part of the algorithm is that eac h neuron on the output grid has a neigh b orho o d and when eac h neuron up dates itself, the neurons in the

PAGE 88

77 Figure 4{1: Dieren t top ological spaces neigh b orho o d up date themselv es accordingly This is describ ed in man y b o oks as \top ologically ordered map", \top ology preserving map" or \top ographical map". There are man y misconceptions and misnomer here. First, the neural net w ork is not suc h a map but it sim ulates suc h a map b ecause the co domain is a discrete space, namely the rectangular grid. Second, \top ographical map" is a misnomer for \top ological map." Third, what the net w ork sim ulates is a con tin uous map. That is the exact in terpretation of the idea that \if the t w o outputs are in a neigh b orho o d then the t w o inputs are also in a neigh b orho o d." \T op ology preserving" is not guaran teed. In fact this is ob vious when the input domain is t w o dimensional and the output domain is one dimensional. What w e get is a plain lling P eano curv e. As it is w ell kno wn that a plain region cannot b e top ologically equiv alen t to a curv e segmen t and the P eano curv e is a con tin uous map but not a top ological map, or homeomorphism. Kohonen only giv es a pro cedure but do es not giv e an ob jectiv e function that this pro cedure optimize. P eople so on nd an in terpretation of SOFM in the con text

PAGE 89

78 of clustering and dimension reduction of data, with mathematical abstraction. Ritter et al. giv e an ob jectiv e function [ 58 ] E f w g = 1 2 X ik M k ( i; k ) jj k w i jj 2 ; where M k is the mem b ership matrix elemen t, equal to 1 if patter is in cluster k and 0 otherwise. ( i; k ) is the neigh b orho o d function b et w een cluster i and cluster k ( i; k ) = 1 if i = k and falls o with distance jj w i w k jj A t ypical c hoice for ( i; k ) is ( i; k ) = e jj w i w k jj 2 2 2 : The Kohonen pro cedure is actually the clustering pro cess of high dimensional data. It tries to mo del the dimensionalit y reduction. The data p oin ts are in h dimensional Euclidean space but they ma y appro ximately lie on a t w o dimensional manifold. The clusters are constrained with a 2D rectangular grid. During the clustering it is required that neigh b oring cluster cen ters sta y close to eac h other. Th us this pro cedure pro vides an appro ximate discrete 2D patc h whic h is a map from a 2D rectangular region with grid to the 2D manifold em b edded in h dimensional Euclidean space, where the data dw ell. This pro vides an in trinsic co ordination (2D) for the data and it extracts 2D features from the data. Because of this clustering pro cess, the co ordinates are only appro ximate and the co ordinates are discrete (the i; j indices). In the general case, eac h data p oin t has h co ordinate comp onen ts and th us liv e in R h but the set of data only p opulate a l -dimensional submanifold M l of R h Supp ose w e ha v e a k -dimensional grid of cluster cen ters, U R k Only when k = l can w e ha v e a homeomorphism from U to an op en neigh b orho o d of M l One example is that l = 3 and k = 2. If the em b edded manifold is roughly a 2D sheet with some non-negligible thic kness in the third

PAGE 90

79 dimension, the clustering result in an appro ximation in whic h w e represen t this 3D curv ed thic k sheet with a 2D thic k-less sheet. If instead of a thic k sheet, w e ha v e a solid 3D manifold and w e still w an t to use a 2D sheet to appro ximate it, it will result in that the 2D sheet will wrinkle and scram ble so that it tries to ll the 3D space. It is easier to visualize when l = 2 and k = 1 and then w e ha v e a space lling P eano curv e, resulting a con tin uous map but not an homeomorphism. 4.3 T op ological Clustering and Matc hing 4.3.1 Wh y: the Need for T op ology When the t w o shap e dier b y a large deformation, complications do o ccur. Figure 4{2 sho ws the con tours of t w o hand shap es. The hand on the left has the th um b and the fore nger p oin ting out while the hand on the righ t has the fore nger folded. These t w o shap es dier b y a large deformation. W e apply our dieomorphic p oin t matc hing algorithm to these t w o shap es and Figure 4{3 sho ws the clustering of the t w o shap es. While the clustering lo oks prett y go o d, a closer examination of the corresp ondence as sho wn in Figure 4{4 indicates that there are incorrect corresp ondences. W e ha v e discussed in Chapter 1 the issues that there is no w a y to clearly dene what the correct corresp ondence is, with our visual in tuition, w e kno w this is not the corresp ondence w e w an t. This is the case where nearb y p oin ts corresp ond to p oin ts that are not nearb y a violation of top ological prop ert y of the mapping. W e w an t to enforce the constrain t that nearb y p oin ts are mapp ed to nearb y p oin ts and in tro duce the top ology constrain t to the matc hing. This has signicance in t w o scenarios: 1. The problem denition of the shap e matc hing do es not ha v e top ology constrain t and the ob jectiv e function do es not ha v e the top ology term. Ho w ev er, the n umerical pro cedure of solving this problem ma y b e caugh t up in a lo cal minim um, whic h giv es incorrect corresp ondence. In this case,

PAGE 91

80 Figure 4{2: Image con tours of t w o hands in tro ducing the top ological constrain t will help a v oid the lo cal minim um and nd the correct corresp ondence as dened b y the ob jectiv e function. 2. The n umerical pro cedure do es nd the global minim um dened b y the objectiv e function without top ological constrain t. Ho w ev er, the corresp ondence is still not what w e in tended. In this case, the p oin t set as a set of p oin t without an y top ological structure is not sucien t to describ e the underlying shap e. F or example of the t w o hand shap es as in Figure 4{2 w e kno w that the p oin ts lie on a con tour curv e. By adding this top ology requiremen t, w e are actually dening a dieren t problem from that without top ology structure and of course the solutions of the corresp ondence should b e dieren t with and without top ology since they are t w o dieren t problems. 4.3.2 Ho w: Graph T op ology It is clear that the information of the top ological structure of the underlying shap e of the p oin t set helps dene and solv e for the correct corresp ondence. Then it is natural to in tro duce some top ological structure in the set of cluster cen ters and a graph is the easiest w a y to represen t this top ology W e mak e the cluster cen ters the v ertices of the graph and assign edges b et w een the v ertices. An in tuitiv e thinking of the top ology construction on the graph ma y b e that mak e the set of v ertices the supp ort set and mak e the adjacen t v ertices

PAGE 92

81 0 50 100 50 0 50 100 150 200 100 150 200 250 50 0 50 100 150 200 Figure 4{3: Clustering of t w o hands 50 100 150 200 0 50 100 150 Figure 4{4: Corresp ondence

PAGE 93

82 B A U 3 U 2 1 U U A nn rr Figure 4{5: Finite top ology neigh b orho o d. A more careful analysis will sho w that this idea should fail. Since the set of the v ertices is a nite set, w e pro v e a prop ert y of nite top ology Theorem 7. L et ( X ; T ) b e a top olo gic al sp ac e and X is a nite set. T is Hausdor i T is discr ete top olo gy. Pr o of. First w e pro v e if T is discrete top ology then T is Hausdor. The pro of is trivial b ecause in discrete top ology ev ery singleton set of a p oin t is an op en set. F or an y t w o distinct p oin ts A 2 X and B 2 X the disjoin t neigh b orho o d U A and U B are U A = f A g and U B = f B g Next w e pro v e if T is Hausdor then T is discrete top ology W e use pro of b y con tradiction. No w w e assume the con trary namely T is Hausdor but T is a top ology other than discrete top ology Then there m ust exist a p oin t A 2 X suc h that f A g is not an op en set. Let U A b e the smallest op en set that con tains A : U A = U 1 \ U 2 \ ::: \ U k ; where U 1 ; U 2 ; :::; U k are all the op en sets that con tains A By assumption, U A m ust ha v e at least another p oin t dieren t from A W e call this p oin t B as sho wn in

PAGE 94

83 Figure 4{5 No w it is ob vious that there is no op en neigh b orho o d of A that do es not con tain B Hence T is not Hausdor. So the original claim is pro v ed. Since the Euclidean space is a Hausdor space and the shap es w e consider as sub-top ological spaces of the Euclidean space are Hausdor spaces, w e are not in terested in graph top ologies that are not Hausdor. Ho w ev er, from the ab o v e theorem w e kno w that the only Hausdor top ology is discrete top ology The discrete top ology is not go o d here b ecause eac h p oin t is an op en set and eac h p oin t can ha v e a op en set consisting of itself. Eac h p oin t is completely discrete, meaning isolated and disconnected. So the discrete top ology in some sense, is no top ology no \nearness" or top ological structure. The correct approac h is through gr aph r e alization or gr aph emb e dding This is the study of a branc h of graph theory top olo gic al gr aph the ory [ 30 ] In tuitiv ely graph realization or graph em b edding is to think the v ertices of the graph as p oin ts in an Euclidean space and the edges of the graphs as the lines or curv es in the Euclidean space connecting the v ertices. With the graph realization or graph em b edding, the graph G is a sub-top ological space of the shap e S as a top ological space. So if the p oin ts in the graph G are in a neigh b orho o d, then they are also in a neigh b orho o d in the shap e top ological space S In the follo wing of this c hapter, w e dev elop top ological clustering tec hniques for the purp ose of dieomorphic p oin t matc hing. It has man y similarities with Kohonen SOFM but there are also man y dierences. First, the purp ose of SOFM is dimensionalit y reduction while our top ological clustering is for p oin t matc hing. Second, the map in SOFM is lo cal while our top ology is global. SOFM only pro vides a single patc h for one op en neigh b orho o d of a p oin t on the submanifold while our clustering allo ws non-trivial top ologies whic h cannot b e co v ered b y a single patc h. Third, the in teraction in SOFM is b et w een one cluster cen ter and

PAGE 95

84 the data p oin ts in the neigh b oring clusters while the in teraction in our mo del in b et w een one cluster cen ter and the neigh b oring cluster cen ters. 4.4 Ob jectiv e F unction and the Algorithm In order to enforce the principle of that p oin ts in a neigh b orho o d should sta y in the neigh b orho o d, w e add the term N X m =1 N X n =1 G mn jj r m r n jj 2 (4.1) to the ob jectiv e function, where G is the adjacency matrix of the graph of the cluster cen ters of one set. The other set has the similar top ological constrain ts. G mn is 1 if there is an edge b et w een r m and r n and 0 otherwise. The new ob jectiv e function no w is E ( M x ; M y ; r ; s ; v ; ) (4.2) = N 1 X i =1 N X k =1 M x ik jj x i r k jj 2 + 2 2 T N 1 X i =1 N X k =1 M x ik log M x ik + X N X k =1 M y j k jj y j s k jj 2 + 2 2 T N 2 X j =1 N X k =1 M y j k log M y j k + N X k =1 jj s k ( r k ; 1) jj 2 + 2 2 T Z 1 0 Z n jj Lv ( x; t ) jj 2 dxdt + 4 N X m =1 N X n =1 G mn jj r m r n jj 2 + 4 N X q =1 N X q =1 G pq jj s p s q jj 2 : The matrix G is a symmetric matrix. The reason for the factor 1 2 in the last line is that eac h edge in the graph is coun ted t wice in the summation. is a new parameter describing the strength of the links b et w een the cluster cen ters. The algorithm to minimize this energy is v ery similar to the algorithm in tro duced in Chapter 3 Ho w ev er, the up date equations for the cluster cen ters ( 3.19 ) and ( 3.20 ) should b e mo died accordingly:

PAGE 96

85 r k = P N 1 i =1 M x ik x k + s k P Nl =1 R 1 0 l ( t ) K ( l ( t ) ; k ( t )) dt + P Nm =1 G mk r m 1 + P N 1 i =1 M x ik + P Nm =1 G mk ; (4.3) s k = P N 2 j =1 M y j k y j + ( r k ; 1) + P Nm =1 G mk s m 1 + P N 2 j =1 M y j k + P Nm =1 G mk ; 8 k : (4.4) The rule of mo dication is, when up dating the cluster cen ter p ositions for r k consider all other cluster cen ters r m ; m = 1 ; 2 ; :::; N in the graph, whenev er there is an edge from r m to r k w e add r m to the n umerator and w e add to the denominator. The mo dication to the up date of s k in the second set is similar. 4.5 Prescrib ed T op ology In some situations, the class of shap es ha v e the same top ology and the top ology is kno wn and the graph represen tation of the top ology is easy In suc h cases, w e can initialize the graph with the prescrib ed top ology In the 2D case, when w e deal with line con tours, t w o t ypical situations are the op en curv es and closed curv es. W e can use c hain top ology and ring top ology for the graphs. 4.5.1 Chain T op ology W e solv ed the matc hing problem again with the hand shap es with c hain top ology The clustering result is sho wn in Figure 4{6 and this w a y w e nd the correct corresp ondence in Figure 4{7 4.5.2 Ring T op ology W e applied the top ological clustering and matc hing algorithm to the corpus callosum data, with ring top ology Figure 4{8 sho ws the top ological clustering while Figure 4{9 sho ws the corresp ondence. 4.5.3 S 2 T op ology W e kno w the hipp o campus shap es ha v e a S 2 top ology W e initialize the graph as a latitude and longitude grid. Figure 4{11 is the top ological clustering of the

PAGE 97

86 0 50 100 50 0 50 100 150 200 100 150 200 250 50 0 50 100 150 200 Figure 4{6: T op ological clustering and matc hing of t w o hands 50 100 150 200 20 0 20 40 60 80 100 120 140 160 Figure 4{7: Corresp ondence with top ology constrain t

PAGE 98

87 0.4 0.6 0.8 1 1.2 0.2 0 0.2 0.4 0.4 0.6 0.8 1 0.2 0 0.2 0.4 Figure 4{8: T op ological clustering of corpus callosum shap es

PAGE 99

88 0.5 0.6 0.7 0.8 0.9 1 0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Figure 4{9: Corresp ondence in top ological clustering of corpus callosum shap es

PAGE 100

89 292 294 296 192 194 196 4 5 6 7 8 9 Figure 4{10: Sphere top ology rst hipp o campus set and Figure 4{12 is the top ological clustering of the second hipp o campus set. 4.6 Arbitrary T op ology W e can see the limitations with prescrib ed top ology First, the class of shap es ma y ha v e dieren t top ologies. Second, ev en if the class of the shap es ha v e the same top ology the top ology ma y b e unkno wn b efore w e run the matc hing algorithm. Third, ev en if the top ology is kno wn, it ma y b e to o complicated to construct a graph appro ximation and the initialization ma y in v olv e h uman in terv en tion. F orth, the edges of the graph is not truly top ological in a sense they are not indenitely rexible and stretc hable strings. The top ological constrain ts in the ob jectiv e function 4.2 are actually iden tical elastic strings. So w e see the geometrical or metric factor in the constrain ts. Sure w e can adjust the co ecien t

PAGE 101

90 250 300 350 160 180 200 220 240 0 5 10 15 Figure 4{11: T op ological clustering of hipp o campus with S 2 top ology: the rst set

PAGE 102

91 250 300 350 120 140 160 180 200 220 0 5 10 Figure 4{12: T op ological clustering of hipp o campus with S 2 top ology: the second set

PAGE 103

92 for dieren t stiness of the strings, but it is dicult to tune the dieren t relativ e stiness connecting the cluster cen ters. Because this elasticit y the graph ma y ha v e resistance to completely t the shap e. This can b e seen in previous hipp o campus examples. The solution is to nd the top ology of the shap e on the ry First w e cluster one of the data set. W e then nd the N-nearest neigh b ors of eac h cluster cen ter. F rom eac h cluster cen ter, w e dra w a directed edge to eac h of its nearest neigh b ors. This w a y w e ha v e a directed graph. The adjacency matrix is not symmetric in general. W e then symmetrize the adjacency matrix, meaning if w e ha v e an edge from no de i to j but not an edge j to i w e then add an edge from j to i This w a y w e get an undirected graph. W e will use this graph top ology for the clustering. Since in the matc hing problem, w e assume the t w o data sets ha v e the same top ology w e will use the same graph for b oth data sets. W e shrink the graph to the cen troid of the data sets in the initialization and w e con tin ue with the join t clustering and matc hing algorithm. Figure 4{13 is the appro ximate graph top ology of hipp o campus data sets with 4 nearest neigh b ors. Figure 4{14 sho ws the top ological clustering of the rst hipp o campus set and Figure 4{15 of the second hipp o campus set. Figure 4{16 a and b sho ws t w o 2D p oin t shap es. Figure 4{18 is the graph top ology w e ha v e learned using 4 nearest neigh b ors after clustering. Figure 4{18 sho ws the top ological clustering of the t w o shap es.

PAGE 104

93 220 240 260 280 300 320 340 360 160 180 200 220 240 2 4 6 8 10 12 14 Figure 4{13: Graph top ology of hipp o campus shap e through learning

PAGE 105

94 250 300 350 160 180 200 220 240 0 5 10 15 Figure 4{14: T op ological clustering of hipp o campus shap e: the rst p oin t set

PAGE 106

95 250 300 350 120 140 160 180 200 220 0 5 10 Figure 4{15: T op ological clustering of hipp o campus shap e: the second p oin t set 100 0 100 200 300 400 500 0 50 100 150 200 250 300 350 400 450 500 100 0 100 200 300 400 500 0 50 100 150 200 250 300 350 400 450 500 (a) (b) Figure 4{16: Fish shap es

PAGE 107

96 0 100 200 300 400 500 50 100 150 200 250 300 350 400 450 Figure 4{17: Graph top ology for the sh shap e with 4 nearest neigh b ors

PAGE 108

97 0 100 200 300 400 500 200 100 0 100 200 300 400 500 600 700 0 100 200 300 400 100 0 100 200 300 400 500 600 700 Figure 4{18: T op ological clustering and matc hing of the sh shap es

PAGE 109

CHAPTER 5 CONCLUSIONS 5.1 Con tributions The need for a go o d p oin t feature matc hing algorithm arises in v arious application areas of medical image analysis. T o m y kno wledge, this is one of the rst attempts at dieomorphic p oin t matc hing in the circumstances of unkno wn corresp ondence. W e ha v e designed an ob jectiv e function and an algorithm to sim ultaneously nd the b est clustering of t w o p oin t sets and a mapping with the least deformation of space. W e require the space deformation to b e a dieomorphic mapping b ecause it is smo oth and homeomorphic at the same time. The essence of this requiremen t is that in a homeomorphic mapping, neigh b oring p oin ts are mapp ed to neigh b oring p oin ts and the same is true for the in v erse mapping. A dieomorphic mapping can preserv e the features of shap es. The dieomorphism parameterization allo ws us to reco v er large deformations while sim ultaneously ac hieving go o d corresp ondence. W e ha v e demonstrated a join t clustering and dieomorphism algorithm and applied it to 2D corpus callosum shap es and 3D hipp o campal p oin t sets in medical imaging. After dieomorphism estimation, the shap e distance, dened as the geo desic distance on the shap e manifold is computed. Since the p oin t sets ha v e dieren t cardinalities and since the shap e distance is only dened w.r.t. the cluster cen ters, w e also computed a mo died Hausdor distance b et w een one original p oin t set and the after-image of a second p oin t set. W e conclude that when the n um b er of cluster cen ters increases, the mo died Hausdor distance decreases. In the pro cess of careful v alidation, w e in v estigated the role of the dieren t n um b ers of clusters using Jensen-Shannon div ergence in the join t clustering and dieomorphism optimization pro cess. 98

PAGE 110

99 W e further prop ose to use a graph represen tation for the shap e top ology information. Results are giv en for prescrib ed top ologies lik e c hain top ology ring top ology { whic h are v ery common in dealing with 2D con tour shap es { and gen us zero closed surface top ology in 3D. W e also in v estigate the top ology problem in general and the learning of top ology with a nearest neigh b or graph. 5.2 F uture W ork In the curren t form ulation, w e still ha v e a free parameter whose v alue has to b e determined. The immediate future goal is to further address (theoretically and exp erimen tally), the role of free parameters. The same framew ork can b e used for atlas estimation. F urthermore, once w e ha v e a turnk ey 3D dieomorphic feature matc hing algorithm, w e plan to use it for hipp o campal shap e classication of epilepsy patien ts [ 69 ]. W e realize with top ological clustering and matc hing, it is a compromise and comp etition b et w een clustering and top ology preserv ation, in the future, w e w an t to separate the t w o phases of corresp ondence and dieomorphism and hop e that ma y further impro v e the matc hing accuracy

PAGE 111

REFERENCES [1] P S. Alexandro. Zur b egr undung der n-dimensionalen top ologie. Math. A nn. 94:296{308, 1925. [2] S. Belongie, J. Malik, and J. Puzic ha. Shap e matc hing and ob ject recognition using shap e con texts. IEEE T r ans. Patt. A nal. Mach. Intel l. 24(4):509{522, 2002. [3] D. P Bertsek as and J. N. Tsitsiklis. Par al lel and Distribute d Computation: Numeric al Metho ds Pren tice-Hall, Englew o o d Clis, NJ, 1989. [4] F. L. Bo okstein. Principal w arps: Thin-plate splines and the decomp osition of deformations. IEEE T r ans. Patt. A nal. Mach. Intel l. 11(6):567{585, June 1989. [5] F. L. Bo okstein. Morphometric to ols for landmark data: Ge ometry and biolo gy Cam bridge Univ ersit y Press, 1991. [6] N. Bourbaki. El ements de Math ematique, III. T op olo gie G en er ale Hermann, 1940-1949; new edition, 1970-1974. English translation, Gener al T op olo gy Addison-W esley Reading, MA, 1966. [7] V. Camion and L. Y ounes. Geo desic in terp olating splines. In Ener gy Minimization Metho ds for Computer Vision and Pattern R e c o gnition pages 513{527. Springer, New Y ork, 2001. [8] S. S. Chern. Dier entiable manifolds Instituto de F sica e Mathem atica, Univ ersidade do Recife, 1959. [9] S. S. Chern, W. H. Chen, and K. S. Lam. L e ctur es on Dier ential Ge ometry W orld Scien tic, Singap ore, 2000. [10] D. Chillingw orth. Dier ential T op olo gy with a View to Applic ations F earon, London, 1976. [11] G. Christensen. Consisten t linear-elastic transformations for image matc hing. In Pr o c e e dings of Information Pr o c essing in Me dic al Imaging|IPMI 99 pages 224{237. Springer-V erlag, Berlin, 1999. [12] H. Ch ui, J. Ram b o, J. Duncan, R. Sc h ultz, and A. Rangara jan. Registration of cortical anatomical structures via robust 3D p oin t matc hing. In Pr o c e e dings of Information Pr o c essing in Me dic al Imaging{IPMI 99 pages 168{181. Springer-V erlag, Berlin, 1999. 100

PAGE 112

101 [13] H. Ch ui and A. Rangara jan. A new algorithm for non-rigid p oin t matc hing. In Pr o c e e dings of IEEE Conf. on Computer Vision and Pattern R e c o gnition{ CVPR 2000 v olume 2, pages 44{51. IEEE Press, New Y ork, 2000. [14] H. Ch ui and A. Rangara jan. A new feature registration framew ork using mixture mo dels. In IEEE Workshop on Mathematic al Metho ds in Biome dic al Image A nalysis{MMBIA 2000 IEEE Press, New Y ork, 2000. [15] H. Ch ui and A. Rangara jan. Learning an atlas from unlab eled p oin t sets. In IEEE Workshop on Mathematic al Metho ds in Biome dic al Image A nalysis (MMBIA) pages 179{186. IEEE Press, New Y ork, 2001. [16] H. Ch ui and A. Rangara jan. A new p oin t matc hing algorithm for non-rigid registration. Computer Vision and Image Understanding 89:114{141, 2003. [17] H. Ch ui, L. Win, J. Duncan, R. Sc h ultz, and A. Rangara jan. A unied nonrigid feature registration metho d for brain mapping. Me dic al Image A nalysis 7:112{130, 2003. [18] T. Co otes, C. T a ylor, D. Co op er, and J. Graham. Activ e shap e mo dels: Their training and application. Computer Vision and Image Understanding 61(1):38{59, 1995. [19] R.H. Da vies, C. Twining, T.F. Co otes, and C.J. T a ylor. An information theoretic approac h to statistical shap e mo delling. In Eur op e an Confer enc e on Computer Vision (ECCV) Lecture Notes in Computer Science, LNCS 2351, pages I I I: 3{20. Springer, Berlin, 2002. [20] M. P Dubuisson and A. K. Jain. A mo died hausdor distance for ob ject matc hing. ICPR94 pages A:566{568, 1994. [21] J. Duc hon. F onctions splines et ve cteurs ale atoir es T ec h. Rep ort 213, Seminaire d'Analyse Numerique, Univ ersite Scien tique et Medicale, Grenoble, 1975. [22] J. Duc hon. F onctions-spline et esp erances conditionnelles de c hamps gaussiens. A nn. Sci. Univ. Clermont F err and II Math. 14:19{27, 1976. [23] J. Duc hon. Splines minimizing rotation-in v arian t semi-norms in sob olev spaces. In Constructive The ory of F unctions of Sever al V ariables pages 85{100, Springer-V erlag, Berlin, 1977. [24] N. Duta, A. K. Jain, and M.P Dubuisson-Jolly Learning 2D shap e mo dels. In IEEE Conf. on Computer Vision and Pattern R e c o gnition (CVPR) v olume 2, pages 8{14, 1999. [25] D. M. Endres and J. E. Sc hindelin. A new metric for probabilit y distributions. IEEE T r ansaction on Information The ory 49:1858{1860, 2003.

PAGE 113

102 [26] R. A. Fisher. On the mathematical foundations of theoretical statistics. Philosophic al T r ansactions of the R oyal So ciety A 222:309{368, 1922. [27] J. Glaunes, A. T rouv e, and L. Y ounes. Dieomorphic matc hing of distributions: A new approac h for unlab elled p oin t sets and sub-manifolds matc hing. In IEEE Computer So ciety Confer enc e on Computer Vision and Pattern R e c o gnition (CVPR 04) v olume 2, pages 712{718, 2004. [28] U. Grenander P Dupuis and M. I. Miller. V ariational problems on ro ws of dieomorphisms for image matc hing. Quarterly of Applie d Math. 56:587{600, 1998. [29] E. Grimson, T. Lozano-P erez, W. W ells I I I, G. Ettinger, S. J. White, and R. Kikinis. An automatic registration metho d for frameless stereotaxy image guided surgery and enhanced realit y visualization. In IEEE Conf. on Computer Vision and Pattern R e c o gnition (CVPR) pages 430{436. IEEE Press, New Y ork, 1994. [30] J. L. Gross and T. W. T uc k er. T op olo gic al Gr aph The ory John Wiley & Sons, New Y ork, 2001. [31] V. Guillemin and A. P ollac k. Dier ential T op olo gy Pren tice Hall, Englew o o d Clis, NJ, 1974. [32] H. Guo, A. Rangara jan, S. Joshi, and L. Y ounes. A new join t clustering and dieomorphism estimation algorithm for non-rigid shap e mathcing. IEEE Workshop on A rticulate d and Non-rigid Motion (ANM) 2004. [33] H. Guo, A. Rangara jan, S. Joshi, and L. Y ounes. Non-rigid registration of shap es via dieomorphic p oin t matc hing. ISBI 2004 2004. [34] R. Hatha w a y Another in terpretation of the EM algorithm for mixture distributions. Statistics and Pr ob ability L etters 4:53{56, 1986. [35] F. Hausdor. Gr undzuge der Mengenlehr e Leipzig, 1914. [36] F. Hausdor. Set The ory T ranslated b y J. R. Aumann. Chelsea, New Y ork, 1957. [37] D. P Huttenlo c her, G. A. Klanderman, and W. J. Ruc klidge. Comparing images using the Hausdor distance. IEEE T r ans. Patt. A nal. Mach. Intel l. 15(9):850{863, Sep. 1993. [38] H. J. Johnson and G. E. Christensen. Consisten t landmark and in tensit ybased image registration. IEEE T r ansactions on Me dic al Imaging 21:450{469, 2002. [39] S. Joshi and M. Miller. Landmark matc hing via large deformation dieomorphisms. IEEE T r ans. Image Pr o c essing 9:1357{1370, 2000.

PAGE 114

103 [40] T. Kohonen. Clustering, taxonom y and top ological maps of patterns. In Pr o c e e dings of the 6th International Confer enc e on Pattern R e c o gnition pages 114{128, Munic h, German y 1982. [41] T. Kohonen. Self-organized formation of top ologically correct feature maps. Biolo gic al Cyb ernetics 43:59{69, 1982. [42] T. Kohonen. Self-Or ganization and Asso ciative Memory Springer-V erlag, 3rd edition, New Y ork, 1989. [43] J. J. Koso wsky and A. L. Y uille. The in visible hand algorithm: Solving the assignmen t problem with statistical ph ysics. Neur al Networks 7(3):477{490, 1994. [44] K. Kurato wski. Sur l'op eration A de l'analysis situs. F undamenta Math. 3:182{199, 1922. [45] C. P Lu and E. Mjolsness. Tw o-dimensional ob ject lo calization b y coarse-tone correlation matc hing. In J. Co w an, G. T esauro, and J. Alsp ector, editors, A dvanc es in Neur al Information Pr o c essing Systems 6 pages 985{992. Morgan Kaufmann, San F rancisco, CA, 1994. [46] S. Marsland C. Twining and C. T a ylor. Measuring geo desic distances on the space of b ounded dieomorphisms. In BMV C 2002 pages 847{856, 2002. [47] Mathematical So ciet y of Japan. Encyclop e dic Dictionary of Mathematics T ranslated b y the Massac h usetts Institute of T ec hnology the MIT press, Cam bridge, 1986. [48] J. Meinguet. Multiv ariate in terp olation at arbitrary p oin ts made simple. J. Appl. Math. Phys. (ZAMP) 30:292{304, 1979. [49] M. Miller, S. Joshi, and G. Christensen. Br ain Warping pages 131{155. Academic Press, New Y ork, 1998. [50] J. W. Milnor. T op olo gy fr om the dier entiable viewp oint Princeton Univ ersit y Press, Princeton, NJ, 1997. [51] E. Mjolsness and C. Garrett. Algebraic transformations of ob jectiv e functions. Neur al Networks 3:651{669, 1990. [52] J. Mo dersitzki. Numeric al Metho ds for Image R e gistr ation Oxford Univ ersit y Press, New Y ork, NY, 2004. [53] N. P aragios X. Huang and D. Metaxas. Establishing lo cal corresp ondences to w ards compact represen tations of anatomical structures. In Pr o c. of the 6th A nnual International Conf. on Me dic al Image Computing and Computer Assiste d Intervention, MICCAI03 pages 926{934, Mon treal, Canada, No v. 2003.

PAGE 115

104 [54] A. Rangara jan, H. Ch ui, and F. Bo okstein. The softassign Pro crustes matc hing algorithm. In Information Pr o c essing in Me dic al Imaging (IPMI '97) pages 29{42. Springer, New Y ork, 1997. [55] A. Rangara jan, H. Ch ui, and J. Duncan. Rigid p oin t feature registration using m utual information. Me dic al Image A nalysis 1999. [56] A. Rangara jan, H. Ch ui, E. Mjolsness, S. P appu, L. Da v ac hi, P GoldmanRakic, and J. Duncan. A robust p oin t matc hing algorithm for autoradiograph alignmen t. Me dic al Image A nalysis 4(1):379{398, 1997. [57] C. R. Rao. Information and the accuracy attainable in the estimation of statistical parameters. Bul l. Calcutta Math. So c. 37:81{89, 1945. [58] H. Ritter and K. Sc h ulten. Kohonen's self-organizing maps: exploring their computational capabilities. In IEEE International Confer enc e on Neur al Networks v olume 1, pages 109{116, San Diego, 1988. [59] K. Rose, E. Gurewitz, and G. F o x. Statistical mec hanics and phase transitions in clustering. Physic al R eview L etters 65(8):945{948, Aug. 1990. [60] A. Rosenfeld and J. L. Plalz. Sequen tial op erations in digital picture pro cessing. Journal of A CM 13:471{494, 1966. [61] B. A. Rosenfeld. History of Non-Euclide an Ge ometry T ranslated b y A. Shenitzer, Springer-V erlag, Berlin, 1988. [62] I. Sc ho en b erg. On in terp olation b y spline functions and its minim um prop erties. Internat. Ser. Numer. A nal. 5:109{129, 1964. [63] I. Sc ho en b erg. Spline functions and the problem of graduation. Pr o c. Nat. A c ad. Sci. U.S.A. 52:947{950, 1964. [64] T. B. Sebastian, J. J. Crisco, P N. Klein, and B. B. Kimia. Construction of 2D curv e atlases. In IEEE Workshop on Mathematic al Metho ds in Biome dic al Image A nalysis{MMBIA 2000 pages 70{77, June 2000. [65] R. Sinkhorn. A relationship b et w een arbitrary p ositiv e matrices and doubly sto c hastic matrices. A nn. Math. Statist. 35:876{879, 1964. [66] C. Small. The Statistic al The ory of Shap e Springer, New Y ork, NY, 1996. [67] H. T agare, D. O'Shea, and A. Rangara jan. A geometric criterion for shap e based non-rigid corresp ondence. In Fifth Intl. Conf. Computer Vision (ICCV) pages 434{439, 1995. [68] H. D. T agare, D. O'shea, and D. Groisser. Non-rigid shap e comparison of plane curv es in images. Journal of Mathematic al Imaging and Vision 16:57{68, 2002.

PAGE 116

105 [69] N. V ohra, B. C. V em uri, A. Rangara jan, R. L. Gilmore, S. N. Rop er, and C. M. Leonard. Kernel Fisher for shap e based classication in epilepsy In Me dic al Image Computing and Computer-Assiste d Intervention (MICCAI02) pages 436{443, 2002. [70] G. W ah ba. Spline Mo dels for Observational Data SIAM, Philadelphia, P A, 1990.

PAGE 117

BIOGRAPHICAL SKETCH Hongyu Guo w as b orn in China. He receiv ed BS and MS degree in ph ysics in Nank ai Univ erisit y He receiv ed his MS degree in computer sciece from the Univ ersit y of Florida. Hongyu Guo has taugh t in a univ ersit y He has also w ork ed in soft w are industry in San F rancisco, CA. His curren t researc h in terests are in computer vision and biomedical imaging. 106


Permanent Link: http://ufdc.ufl.edu/UFE0011645/00001

Material Information

Title: Diffeomorphic Point Matching with Applications in Medical Image Analysis
Physical Description: Mixed Material
Copyright Date: 2008

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0011645:00001

Permanent Link: http://ufdc.ufl.edu/UFE0011645/00001

Material Information

Title: Diffeomorphic Point Matching with Applications in Medical Image Analysis
Physical Description: Mixed Material
Copyright Date: 2008

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0011645:00001


This item has the following downloads:


Full Text











DIFFEOMORPHIC POINT MATCHING
WITH APPLICATIONS IN MEDICAL IMAGE ANALYSIS















By

HONGYU GUO


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA


2005

































Copyright 2005

by

Hongyu Guo















To my daughter Alicia.















ACKNOWLEDGMENTS

I will be forever indebted to my advisor, Dr. Anand Raingirajain, for his

invaluable guidance and support. I am very lucky to have the opportunity to

work with him. I am very grateful for his appreciation and confidence in my

mathematical talent. Although he is an outstanding researcher in the computer

vision and medical imaging field with many distinguished contributions, he is very

humble and down to earth. During the years of my study, I could knock at his door

any time I needed him and he would put down his work and discuss sparkling ideas

with me. Working with him has been truly a wonderful and rewarding experience.

I would like to express deepest thanks to Dr. Arunava Ban'. i Dr. Jorg

Peters and Dr. Baba Vemuri of the Department of CISE, Dr. Yunmei Chen of the

Department of Mathematics and Dr. Haldun Aytug of the Department of Decision

and Information Sciences for serving on my supervisory committee. I thank them

for their invaluable criticism and advice.

I wish to extend my gratitude to Dr. Sarang Joshi at the University of North

Carolina at Chapel Hill and Dr. Laurent Younes at Johns Hopkins University.

They are the coauthors of a book chapter and several papers that I have written

and this work in my dissertation is greatly benefitted from their previous work on

diffeomorphic landmark matching.

I would like to thank Dr. Eric Grimson, the director of MIT AI Lab, for the

helpful discussion and valuable comments when he visited the University of Florida.

I wish to extend special thanks to Dr. Alper for an enlightening discussion on

triangulation and surface reconstruction.









My thanks also go to ev, i,..i i. of my fellow students at the Center for Com-

puter Vision, Graphic sand Medical Imaging at the University of Florida, particu-

larly, Bing Jian, Santhosh Kodipaka, Jie Z1i.,r-. Adrian Peter, Zhizhou Wang, Eric

Spellman, Fei Wang, Tim McGraw and Xiaobin Wu. I get constant help from them

every day.

I am indebted to Dr. John M. Sullivan at the University of Illinois at Urbana-

Champaign for granting permission to use his graphics image of the Klein bottle.

I would like to thank my wife Yanping for her love and support and I thank

my daughter Alicia for being a wonderful daughter. Alicia and I enjoyed working

together on many weekends. While she was fascinated by the rotating shapes of

computer graphics I displayed with MATLAB, I admired her beautiful and creative

colorings and drawings.

This research is supported in part by the National Science Foundation grant

IIS 0307712.















TABLE OF CONTENTS
page

ACKNOWLEDGMENTS ................... ...... iv

LIST OF TABLES ................... .......... viii

LIST OF FIGURES ..................... .......... ix

ABSTRACT .................................... xi

CHAPTER

1 INTRODUCTION ................... ......... 1

2 PREVIOUS WORK ................... ....... 17

2.1 Thin Plate Splines (TPS) .......... .. ........... 18
2.2 Bookstein's Application to 2D Landmark Warping ........ 19
2.3 More on Splines .................. .......... 21
2.4 Reproducing Kernel Hilbert Space (RKHS) Formulation ...... 23
2.5 The Folding Problem of Splines ...... ........... 24
2.6 Imposing Restriction on the Jacobian ....... ....... 25
2.7 The Flow Approach ................... .... 25
2.8 Correspondence and Softassign ........... ... .. 28
2.9 Distance Transforms .............. ... .. 28
2.10 Implicit Correspondence ..... ........... . .. 29
2.11 Shape Context ............... ......... .. 31
2.12 Active Shape Models . . . ..... ..... 31
2.13 Deterministic Annealing Applied to EM Clustering . ... 32
2.14 Statistical Shape Analysis on Differentiable Manifolds ...... ..34
2.15 Distance Measures from Information Theory . . ... 36

3 DIFFEOMORPHIC POINT MATCHING ................ .. 42

3.1 Existence of a Diffeomorphic Mapping in Landmark Matching 42
3.2 Symmetric Matching due to Time Reversibility . .... 44
3.3 A Theoretical Framework for Diffeomorphic Point Matching .. 46
3.4 A Diffeomorphic Point Matching Algorithm . . .. 49
3.5 Applications to 2D Corpus Callosum Shapes . . ... 53
3.6 Applications to 3D Shapes ................. . .. 55
3.6.1 Experiments on Synthetic Data . . 59
3.6.2 Experiments on Real Data .... . . 61









4 TOPOLOGICAL CLUSTERING AND MATCHING .......... 72

4.1 Fundamentals of Topological Spaces ...... ........... 72
4.2 Kohonen Self-Organizing Feature Map (SOFM) . ... 76
4.3 Topological Clustering and Matching . . ..... 79
4.3.1 Why: the Need for Topology . . . 79
4.3.2 How: Graph Topology ............... . .. 80
4.4 Objective Function and the Algorithm . . 84
4.5 Prescribed Topology .................. ..... .. 85
4.5.1 Chain Topology ...... ....... .. .. .. 85
4.5.2 Ring Topology ................ ... .. 85
4.5.3 S2 Topology ............... ..... .. 85
4.6 Arbitrary Topology ............... ...... 89

5 CONCLUSIONS ............... .......... .. 98

5.1 Contributions ............... ......... .. 98
5.2 Future W ork ................. ...... ...... 99

REFERENCES ................... .......... .. .. 100

BIOGRAPHICAL SKETCH .................. ......... .. 106















LIST OF TABLES
Table page

3-1 Modified Hausdorff distance of the matching point sets. . ... 57

3-2 Geodesic distances and modified Hausforff distances . ... 58

3-3 Matching errors on synthetic data with different noise levels . 60

3-4 Limiting value of a determined by the number of clusters ...... ..61

3-5 Jensen-Shannon divergence for various pairs of shapes . ... 70

3-6 Hausdorff and modified Hausdorff distance for various pairs of shapes 71















LIST OF FIGURES
Figure page

1-1 Grids in R2 (a) before transformation and (b) after transformation .. 9

1-2 Klein bottle immersed in R3. ............. .... .. 9

1-3 (a) The template image and (b) The reference image . ... 10

1-4 Space deformations .................. ......... .. 10

1-5 Ambiguity of landmark correspondence ................ .. 13

1-6 Two shapes of a ribbon and the associated landmarks . ... 15

1-7 Misinterpretation of the shape .................. ..... 16

2-1 Deformation of the thin plate .................. ..... 38

2-2 Landmark displacements .................. ..... .. 38

2-3 Thin-plate Spline interpolation .................. ..... 39

2-4 Diffeomorphic interpolation .. ......... .... 39

2-5 Two distance transformed images of three landmarks . ... 40

2-6 Level sets of two distance transformed images . . 41

3-1 Existence of a diffeomorphic mapping . . ...... 44

3-2 Asymmetry of the matching ................ .... 45

3-3 Point sets of nine corpus callosum images. . . 56

3-4 Clustering of the two point sets. ............. . 56

3-5 Diffeomorphic mapping of the space. ................. 56

3-6 Matching between the two point sets. ................. 57

3-7 Overlay of the after-images of eight point sets with the ninth set 57

3-8 Two point sets of hippocampal shapes ... . ... 59

3-9 Matching errors on synthetic data for different number of clusters .65

3-10 Two hippocampal shapes ................ ...... 66









3-11 Deterministic annealing in the clustering process . . .... 67

3-12 Limiting value of a determined by the number of clusters ...... ..68

3-13 Clustering of the two hippocampal shapes . . 69

4-1 Different topological spaces ................ ..... 77

4-2 Image contours of two hands ............... ... 80

4-3 Clustering of two hands ............... .... .. 81

4-4 Correspondence ............... ........... .. 81

4-5 Finite topology ............... ........... .. 82

4-6 Topological clustering and matching of two hands . . ... 86

4-7 Correspondence with topology constraint . . ...... 86

4-8 Topological clustering of corpus callosum shapes . . ... 87

4-9 Correspondence in topological clustering of corpus callosum shapes .. 88

4-10 Sphere topology ............... ........... .. 89

4-11 Topological clustering of hippocampus with S2 topology: the first set 90

4-12 Topological clustering of hippocampus with S2 topology: the second set 91

4-13 Graph topology of hippocampus shape through learning . ... 93

4-14 Topological clustering of hippocampus shape: the first point set 94

4-15 Topological clustering of hippocampus shape: the second point set 95

4-16 Fish shapes .................. ............. .. 95

4-17 Graph topology for the fish shape with 4 nearest neighbors . 96

4-18 Topological clustering and matching of the fish shapes . ... 97















Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

DIFFEOMORPHIC POINT MATCHING
WITH APPLICATIONS IN MEDICAL IMAGE ANALYSIS

By

Hongyu Guo

August 2005

Chair: Dr. Anand Rangarajan
M., i r Department: Computer and Information Science and Engineering

Diffeomorphic matching of unlabeled point sets is very important to non-rigid

registration and many other applications but it has never been done before. It is a

very challenging problem because we have to solve for the unknown correspondence

between the two point sets. In this work we propose a joint clustering method to

solve for a simultaneous estimation of the correspondence and the diffeomorphism

in space. The cluster centers in each point set are always in correspondence by

virtue of having the same index. During clustering, the cluster center counterparts

in each point set are linked by a diffeomorphism and hence are forced to move

in lock-step with one another. We devise an objective function and design an

algorithm to find the minimizer of the objective function. We apply the algorithm

to 2D and 3D shapes in medical imaging. We further propose to use a graph

representation for the shape topology information. Results are given for prescribed

topologies like chain topology, ring topology -which are very common in dealing

with 2D contour shapes -and genus zero closed surface topology in 3D. We also

investigate the topology problem in general and the learning of topology with a

nearest neighbor graph.















CHAPTER 1
INTRODUCTION

This work is a first time effort to study the diffeomorphic point matching

problem with unknown correspondence. In this chapter we introduce the need for

an effective diffeomorphic point matching algorithm and the background knowledge

of the shape matching field. We give some fundamental definitions and theorems

from differential topology which are essential in the development of the following

chapters. We analyze the nature and difficulty of the correspondence problem.

Shape analysis using tools of modern differential geometry and statistical

theory has far and wide applications in computer vision, image processing, biology,

morphometrics, computational anatomy, biomedical i,,,..2-i,,. and image guided

surgery as well as archeology and astronomy. Shapes play a fundamental role in

computer vision and image analysis and understanding. In general terms, the shape

of an object, a data set, or an image can be defined as the total of all information

that is invariant under certain spatial transformations [ ]. Shape matching and

correspondence problems arise in various application areas such as computer vision,

pattern recognition, machine learning and especially in computational anatomy

and biomedical imaging. Shape matching becomes an indispensable part of many

biomedical applications like medical diagnosis, radiological treatment, treatment

evaluation, surgical planning, image guided surgery and pathology research.

Shapes may have many different representations. They can be represented

with the intensities of pixels of an image, which is a function defined in a region

of 2D or 3D space, or they can be represented with point sets, curves or surfaces.

In this dissertation, we focus on the point representation of shapes. Point repre-

sentation of image data is widely used in all areas and there is a huge amount of









point image data acquired in various modalities, including optical, MRI, computed

tomography and diffusion tensor images [ ].

The advantage of point set representation of shapes, as opposed to curve and

surface representations is multifarious and that is why we focus on point shape

matching, instead of curve or surface matching [ ]. The point set repre-

sentation is compact in storage. The computational time when using point set

representation is dramatically reduced as opposed to when image intensities are

used. The point set representation of shapes is universal and homogeneous. It does

not require the prior knowledge about the topology of shapes. It has the capability

to fuse different types of features into a global, uniform and homogeneous repre-

sentation. A point set representation of shapes is especially useful when feature

grouping (into curves and the like) cannot be assumed. Statistical analysis on point

set shapes is straightforward, as demonstrated in Cootes et al. [ ] using active

shape models. The recent work of Glaunes et al. [ ] is a bold step forward to

generalize the point set shapes to general Radon measures and distributions in

the sense of Schwartz generalized functions, in order to model shapes represented

by a mixture of points and submanifolds of different dimensions (curves and sur-

faces). Although this is viable in theory, it becomes unpractical when it comes to

applications. In their experiments dealing with a mixture of points and a curve,

they used the technique of resampling the curve. That comes back to the point set

representation of the curve itself. Moreover, by doing so, new free parameters, i.e.,

the relative weight of measure between points and the curves, is introduced, which

can be arbitrary.

Point shape matching is ubiquitous in medical imaging and in particular, there

is a real need for a turnkey, non-rigid point shape matching algorithm [ ].

Point shape matching in general is a difficult problem because, as with many

other problems in computer vision, like image registration and segmentation, it









is often ill-posed. We try to make abstractions out of the practical problems to

formulate precise mathematical models of the problem. Point matching can be

viewed in the context, or out of the context of image registration. In the context

of image registration, the points are viewed as feature points in the image. The

correspondence can be known or unknown. When the cardinality of the two point

sets is the same and when the correspondence is known, we call this the landmark

matching problem. The number of feature points may also be unequal. In that

case we are dealing with outliers. The difficulty increases dramatically when the

correspondence is not known and/or when there are outliers. The point matching

problem can also exist out of the context of image registration. In this case, the

points are samples from a shape and we have a point representation of the shape.

When we have two such shapes represented by points, usually the number of points

in the two shapes is different and there is no point-wise correspondence. We want

to find the correspondence between the two shapes. We call this the point shape

matching problem. In the following, we will first address the landmark matching

problem and then the point shape matching problem.

We assume the image domain is d-dimensional Euclidean space Rd. Usually

d = 2 or d = 3. Suppose we have two images, II : 1 --- R and 2 : 2 -- R,

where Q1 C Rd and 22 c Rd. The image registration problems can be classified

into two categories: intensity based registration and feature based registration [ ].

In the intensity based image registration, we need to find a map f : Q1 2 such

that Vx E Q1, Ii(x) = I2(f(x)). This is the ideal registration problem. In feature

based registration, we suppose we have two corresponding sets of feature points, or

landmarks, {pi E 1| i = 1, 2,..., n} and {qi E Q21 i = 1, 2,..., n}. We need to find a

transformation f : Q1 2 such that Vi = 1, 2, ..., n, f(pi) = qi.

In many applications, we are required to find the transformation within some

restricted groups, like rigid transformations, similarity transformations, affine









transformations, projective transformations, polynomial transformations, B-spline

transformations and i' '11- ii.1l transformations. Different transformation groups

have different degrees of freedom, the number of parameters needed to describe a

transformation in the group. This also determines the number of landmark pairs

that the transformation can exactly interpolate. Let us look at some examples. In

the two dimensional space, where d = 2, a rigid transformation, which preserves

Euclidean distance, defined by 1.1 has 3 degrees of freedom (Q, xo, yo) and cannot

interpolate arbitrary landmark pairs.


x' cos O sin x Xo
+ (1.1)
y' sin cos y yo

The landmark pairs to be matched must be subject to some constraints. That is,

they have to have the same Euclidean distance. A similarity transformation defined

by 1.2 has 4 degrees of freedom (k, ;, xo, yo) and can map any 2 points to any 2

points.



=k + (1.2)
y' /sin cos y yo

An affine transformation defined by 1.3 has 6 degrees of freedom

(all, a12, a21, a22, X0, yo) and can map any 3 non-degenerate points to any 3 non-
degenerate points.


X/ all a12 X Xo
+ (1.3)
a21 a22 y yo

A projective transformation defined by 1.4 in term of non-homogeneous

coordinates has 8 degrees of freedom and can map any 4 non-degenerate points to

any 4 non-degenerate points.












/ allx+a12y+al3
dllx+dl2y+l (1.4)
y! [ bll+bl2y+b3i
Sdllx+dl2y+l

In three dimensional space, where d = 3, we write the transformation in a

more general form,



x all a12 a13 x

y' 21 a22 a23 y + Y (1.5)

Z a31 a32 a33 z Z0

where



all a12 a13

A a21 a22 a23

a31 a32 a33

is a 3 x 3 matrix, which represents a linear transformation. A rigid transformation

requires A to be orthogonal and has 6 degrees of freedom. A similarity transfor-

mation requires A to be a similar matrix and has 7 degrees of freedom. An affine

transformation allows A to be the most general form and has 12 degrees of freedom

and can map any 4 non-degenerate points to any 4 non-degenerate points. A pro-

jective transformation in 3D as defined by 1.6 has 15 degrees of freedom and can

map any 5 non-degenerate points to any 5 non-degenerate points.


I allx+a12y+al3z+a14
dllx+d+dl213z+l
y/ Ix bli+bzy+b1z+b4 (1.6)
dlix+dl2+dl2 y+d 1
z c 11 +c12 y+c13 Z+C14
dllx+d+dl213z+l
The term II.I .-irigid" transformation is often used in a narrower sense. Al-

though similarity, affine and projective transformations do not preserve Euclidean

distance, they all have finite degrees of freedom. In the literature, w'1 .i-rigid"









transformation usually refers to a transformation of infinite degrees of freedom,

which can potentially map any finite number of points to the same number of

points. So we immediately see a big difference between finite degree of freedom

transformations and non-rigid transformations. Given a fixed number of landmark

pairs to be interpolated, the former is easily over constrained; the latter is always

under constrained. This is one of the reasons that the non-rigid point matching

problem is much more difficult. To find a unique non-rigid transformation, we need

further constraints. We call this regularization.

Two desirable properties of non-rigid transformations are smoothness and

topology preserving. Let R1 C Rd and Q2 c Rd. A transformation f : Q1 --- 2

is said to be smooth if all partial derivatives of f, up to certain orders, exist and

are continuous. If transformation f : Q >1 -- 2 preserves the topology, then Q1

and Img(f) = {j_ E 13'1 E Q1, '_ = f(pi)} have the same topology. A

transformation that preserves topology means we will not have tears in space or in

the image. A transformation that preserves topology is called a homeomorphism

and its formal definition is:

Definition 1. Let Q1 and Q2 be two /,. .I..,/', ,i spaces. A map f : R ~2 is a

homeomorphism if

f is a bijection;

f is continuous;

the inverse f-1 is continuous.

A smooth transformation f : Q1 Q2 may not preserve the topology. Namely,

a smooth map may not be a homeomorphism. It is easy to see this because a

smooth map even may not be a bijection. On the other hand, a homeomorphism

may not be smooth because in the definition, we only require continuity in both f

and its inverse but we do not require differentiability. It is strongly desirable that









the transformation is both smooth and topology preserving. What we want is a

diffeomorphism which is defined as follows.

Definition 2. Let M1 and 1i be differentiable ri.,,;'-,1-ll. A map f : M1 -> is

a diffeomorphism if

f is a bijection;

f is differentiable;

the inverse f-1 is differentiable.

Because the concept of diffeomorphism is essential to this work, we would

have a little more discussion here in order to clarify some common misconceptions

about this concept. It is obvious that a diffeomorphism is both a smooth map and

a homeomorphism from the definitions. However, what is not so obvious is that by

requiring a diffeomorphism, we are asking for more than something that is both

smooth and homeomorphism. Let us look at some of following counterexamples

and we will learn what factors may contribute to make the transformation fail to be

a diffeomorphism.

First, a smooth bijection is not necessarily a diffeomorphism. Let M1 C R

be [0, 27) and M2 = S1 the unit circle. f : M -1 M2 defined by f(p) =

(cos(P), sin(p)) is smooth and bijective, but not a diffeomorphism because the
inverse f-1 is not continuous, and hence not differentiable. In fact, f is not a

homeomorphism.

Second, a smooth homeomorphism is not necessarily a diffeomorphism.

Consider f : R -- R with f(x) = xA. It is smooth and it is a homeomorphism but

the inverse f-1 is not differentiable at x = 0.

A related concept is local diffeomorphism and it is defined as:

Definition 3. A differentiable map f : M -- M2 is a local diffeomorphism if for

each x E MI there exists a neighborhood U of x such that f u : U f(U) is a

diffeomorphism.









Furthermore, it is useful that we list without proof some known facts in differ-

ential topology [ ]about diffeomorphism and local diffeomorphism.

Theorem 1. A map f : M1 [_ is a local diffeomorphism if and only if its

tangent map is an isomorphism.

With some simple facts in linear algebra, the above theorem can be rewritten

as the following.

Theorem 2. A map f : M1 --+ is a local diffeomorphism if and only if its

Jacobian of the tangent map is nowhere equal to zero.

Theorem 3. A map f : M1 M2 is a diffeomorphism if and only if it is a

bijection and a local diffeomorphism.

With the help of these theorems we can visualize more situations when a

smooth map fails to be diffeomorphism. One situation is when the tangent map

fails to be an isomorphism everywhere. Namely, the Jacobian is zero at some

points. In that case, the smooth map even fails to be a local diffeomorphism. The

other situation is that the smooth map is a local diffeomorphism but not a global

diffeomorphism. The following example demonstrates this.

Consider f : R2 -+ R2 with f(x, y) = (e" cos y, e" sin y). The Jacobian is


_f_ af e cosy -e siny
J= e2 0 V(z, y) R2.
e" sin y e cosy

So f is a local diffeomorphism. However, f is not a diffeomorphism as can be seen

in Figure. 1-1. Notice that because the function is periodic in y, in the codomain,

the image of the function has infinitely many sheets overlaid. Another related

example is the self intersection of the immersion of the Klein bottle in R3 as shown

in Figure. 1-2.

Now let us look at an example of another smooth transformation, namely, the

thin-plate spline (TPS) [ ].


























Figure 1-1: Grids in R2 (a) before transformation and (b) after transformation


Figure 1-2: Klein bottle immersed in R3.
Image courtesy of John Sullivan. (http://torus.math.uiuc.edu/jms/Images/klein.html)









































(a) (b)


Figure 1-3: (a) The template image and (b) The reference image


Figure 1-4: Space deformations (a) obtained with thin-plate spline interpolation.

The folding of space is illustrated by the deformation of the grid lines. (b) The

desired space transformation: a diffeomorphism, which eliminates the space folding

problem.


f7~m
~L~LhU"//"f
fflj ~ff/J
'711


~LLI ///////1LCn
tt///~C*L~CCCC~
r
iit*amLT~
~Yffffff
/I////////
~fffi~









Figure 1 3a shows a template image. Figure 1 3b is the reference image

obtained by warping the template. Some landmarks are selected and shown in

the images. Figure 1 4a demonstrates the transformation of space by showing

the deformation of the rectangular grid. We can see the folding of space. This

is the drawback of the thin-plate spline interpolation. Due to the folding of

space, features in the template may be smeared in the overlapping regions. And

furthermore, the transformation is not invertible. A diffeomorphic transformation is

strongly desirable, which preserves the features, the topology and which is smooth

as shown in Figure 1 4b.

We still face the unknown correspondence problem, which is a difficult

problem. There are two scenarios in which we encounter the correspondence

problem. The first scenario is when we do landmark based image registration. If

the landmarks are selected by hand, we do know the correspondence. However, the

process of hand picking landmarks is painstaking and it requires expert knowledge

about the image and the subject area and it may involve human error. The

complete automation of landmark selection is still not achieved at present but there

is good progress towards the goal. If the landmarks are automatically selected, then

the correspondence is unknown and the correspondence needs to be automatically

obtained. The second scenario is when we have a point set representation of the

shapes. Each shape is represented by a large point set (point cloud) and we have

no knowledge about the correspondence between the two point sets.

What makes the unknown correspondence problem more difficult is that it is

ill-posed. The correspondence between two point sets is a very intuitive concept

which ev- i1-',il seems to understand. However, as is well known within the medical

image analysis community, it is very difficult to define correspondence precisely.

This creates a problem for validation since an irreducibly subjective factor seems to

be present in deciding what is a good correspondence.









We continue with a short discussion aimed at reaching a better understanding

of point correspondence. Suppose we have two point sets S1 { pi E 1I i

1, 2,..., n} andS2 {qi E 2i = 1, 2,..., n} where Q C Rd and Q2 c Rd,

usually with d = 2, or d 3. When S1 and S2 are point sets of equal cardinality

with points randomly distributed in space, what is the correspondence between

S1 and S2? There are n! bijections, or permutations between the two sets. We

have no a priori knowledge allowing us to judge that one correspondence is better

than the other. Finite point sets have no extra structure over and beyond their

discrete structure. So when we talk about the correspondence between point sets,

we always imply that the point set is the representation of some underlying shape,

which is a topological space. And, as is fairly standard in the literature, we can

bypass discussing point correspondence by focusing on the correspondence of the

underlying topological space, which is a homeomorphism, and most desirably a

diffeomorphism. Even so, the homeomorphism between the two topological spaces

is not unique. Here we make our first assumption: the optimal correspondence

between the two point sets is the one that induces the least deformation of space.

Here we understand the notion of space deformation intuitively but we will

carefully define it later. It is easy to realize that this assumption is not always true

because we are only given the point set and we have no explicit knowledge of the

underlying shape.

This is easily illustrated with a simple example. In Figure 1-5a and 1-5b

we have two point sets. It is not clear at all what the correspondence is between

the two point sets. Figure 1-5c and 1-5d put the two point sets on top of two

underlying images, the images of two women. Figure 1-5c is the drawing My Wife

and My Mother-in-law published in 1915 by the cartoonist W. E. Hill. Figure 1-5d

is the image of Head of a Woman in Profile (cropped) by Jean-Pierre David. If

you interpret Figure 1-5c as the image of a young girl, namely the point C as the





























A
B H

*G
C F
D


SM

L
I J
J


50 100 150

(a)









AB


50 100

(b)


150 200


r I
B^

.-'-^ "
< ^.a


(c) (d)

Figure 1-5: Ambiguity of landmark correspondence









chin and H as the ear, then we may have the correspondence between the two point

sets from Figure 1-5a to 1-5b: A-A, B-B, C-C, D-D, E-E, F-F, G-G and

H-H. From another perspective, if we view the image as that of an old woman,

namely point C as the tip of the nose, E the chin and H the left eye, we may have

the correspondence: A-A, B-B, C-I, D-J, E-C, F---, G-L and H-M.

Let us look at another example. Figure 1 6a and Figure 1 6b are two images

of a black strip with the latter non-rigidly deformed. We select four landmarks Al

through A4 for shape A and four landmarks Bi through B4 for shape B. Figure

1 6c shows the landmarks for shape A without the underlying ribbon shape. Figure

1 6d shows the landmarks for shape B without the underlying ribbon shape.

Given the fact that we can see the shapes, we know the correspondences are

Al B1, A2 B2, A3 B3, A4 B4. Figure 1 7a shows this landmark

correspondence and Figure 1 7b shows the deformation of space for this landmark

correspondence. However, if we only have the two sets of points as shown in Figure

1 6c and Figure 1 6d, without the knowledge of the underlying shapes, the above

correspondence is not the one that we will find since it does not give the least

deformation of space. In fact, another correspondence, Al B1, A2 B3,

A3 B2, A4 B4 shown in Figure 1 7d gives a smaller deformation solution as

shown in Figure 1 7e. This is in fact a misinterpretation of the shape, with Figure

1 7c as the original shape and Figure 1 7f the misinterpreted shape, due to the

lack of information regarding the underlying shapes.

Consequently, by making the above assumption of least space deformation,

we are really assuming that the points are dense enough such that the underlying

shape is well represented by the points. Adding more points can help resolve this

ambiguity and hence keep the above assumption-that the correct correspondence

gives the smallest deformation of space-approximately valid.
























Al


(a)
Al

0 -
0 -

A2 *

SA3



0 -A4
0 -
0 -


A4
0 20 40 60 80 100 120 140 160 180 20(


(b)
B1


* B2


B3 *


.B4
100 11
(d)


Figure 1-6: Two shapes of a ribbon and the associated landmarks. (a)
(b) Shape B, (c) Landmarks of shape A without the underlying shape,
marks of shape B without the underlying shape


Shape A,
(d) Land-


i40 i61










Al 1 B1


A2 B2 B2
B3 A3 B3




(a) (b) (c)
Al B1 E B1


A2 B2
B3 A3


A4 ,B4_ t 4
(d) (e) (f)

Figure 1-7: (a) The "correct" correspondence. (b) The space deformation ac-
cording to the "correct" correspondence is not the smallest. (c) The "correctly"
deformed shape. (d) Another possible but 'i'ii, 1,' t" correspondence. (e) The
space deformation with this 'ii ..''i t" correspondence is the smallest. (f) The
misinterpretation of the shape due to the iii .. t" correspondence.


The rest of the dissertation is organized as follows. In Chapter 2 we briefly re-

view the previous related work in the past ten years. In Chapter 3 we describe our

theory of diffeomorphic point matching, develop an algorithm to solve the problem

and apply the algorithm to 2D and 3D medical imaging applications. Chapter 4

is an extension to the work in Chapter 3 and in it we describe a method of topo-

logical clustering and matching. In Chapter 5 we summarize our contributions and

point out the directions of future work.















CHAPTER 2
PREVIOUS WORK

In this chapter, we give a brief account of the previous research that is related

to our work. Most of the work was developed in the past ten years but some work

may be traced back to the 1970s. We start with thin-plate splines, which play

an important role in shape ni.il liii. :. and all other splines work with the same

principle. The Reproducing Kernel Hilbert Space formulation gives us a great

insight on the nature of the entire class of spline warping problems. We then

discuss the folding problem with splines and various approaches to overcome this,

including imposing constraints on the Jacobian and the flow approaches. We also

discuss the correspondence problem. In general there are two classes of approaches

to tackle the correspondence problem. One is to confront it directly while the

other is to try to circumvent it. With the second class of approaches, people have

used implicit correspondence by using a shape distance that is the function of the

two point sets, instead of depending on the point-to-point correspondence, and

there are also some methods that transform the point matching problem to image

matching problem through distance transform. With the first class of approaches,

the direct approach, there are methods that treat the correspondence and matching

separately, like softassign and there are methods that handles both at the same

time like the joint clustering and matching (JCM) algorithms. Shape context

method makes use of extra local context information to help better resolve the

ambiguous point correspondence problem. Active shape models play as a bridge

between the landmark based methods and image intensity based methods, and

also a bridge between rigid matching and non-rigid matching. We spend a section

discussing the deterministic annealing method for EM (1i1-i I, which is an









effective method to avoid local minima and we adopt such an annealing approach

in our algorithm. Also closely related is the work of statistical shape analysis on

differentiable manifold and some distance measures from Fisher information theory.

2.1 Thin Plate Splines (TPS)

A spline is a long strip of wood or metal that is fixed at a number of points.

This has long been used as a drafting tool to draw smooth curves that are required

to pass certain points (called "ducks," or "dogs," or ',.i '). These wooden strips

are not only used in di. flii,-:. but also used in constructions. The Wright brothers

used one to shape their wings. In the old days splines were used in shipbuilding.

They are also used for bending the wood for musical instruments like pianos,

violins, violas, etc. In the modern days, splines are used to model the body of

automobiles. In all those uses, it is actually the ] li-,-i .,1 realization of some smooth

curves. In pilr -i. -. the shape of the wood strip has to take the form of such a curve

that the bending energy of the strip is minimized. Sheoenberg [ ] is credited

to be the first to study spline functions started with one dimensional problems,

the cubic splines. The thin plate spline is the natural generalization of the cubic

spline. In the drafting practice, the wooden strip is long but very thin and narrow,

so that it is abstracted as a one dimensional curve. In the thin plate case, we have

a plate which is very thin, so the problem is two dimensional. In both cases, the

same lir, -i- .,l bending energy of the strip or the thin plate in minimized.

The problem of thin plate spline interpolation is stated as follows. Given

{pi E R21 i 1, 2,..., n} and {zi E R i = 1, 2,..., n}, we want to find a function

f : R2 -- R2 such that it interpolates the points with f(pi) z= i = 1, 2, ...n and
minimizes the thin plate bending energy

E(f) 12 [(f )2 +2( )2] dxdy. (2.1)









If we interpret the {zi E RI i = 1, 2, ..., n} as the displacements of the thin plate in

the z direction at points {pi E R2 i = 1, 2, ..., n} in the x y plane, we can easily

see the 1.r, -i, .,l intuition of this interpolation problem, as shown in Figure. 2-1.

Sometimes it is desirable that we sacrifice the exact interpolation. We only

seek an approximation of the interpolation as a trade off of less deformation of the

thin plate. So the thin plate smoothing problem is defined as given {pi E R2i

1, 2,..., n} and {zi E R| i = 1, 2,..., n}, we want to find a function f : R2 R2 such

that it minimizes the cost function



ii f(p))2 AE(f). (2.2)
i=1

2.2 Bookstein's Application to 2D Landmark Warping

Bookstein [ ] applied thin plate splines to the landmark interpolation

problem. For simplicity, we discuss the problem in 2D space. Everything in the

2D formulation easily applies to 3D except we have a different kernel for 3D. The

goal is to find a smooth transformation f : 1 Q-> 2 that interpolates n pairs of

landmarks {pi E Q1 i = 1, 2, ..., n} and {qi E c 2 i = 1, 2, ..., n}. It is obvious that

such a smooth transformation is not unique. We will give it further constraints.

One choice is to require the to transformation minimize some functional, in this

case



E 2 [,I2L )2 + 2 2 +( )2 dxdy, (2.3)
h= 1
where fi and f2 are the x and y components of the il.,i'1.ii'..



f(x, ) =(fl(x, y), f2(x, y)). (2.4)

If we interpret each of fi and f2 as the bending in the z direction of a metal sheet,

or thin plate, extending in the x y plane, the energy in (2.3) is the analogy of the









thin plate bending energy. The problem can be solved using a standard Green's

function method. The Green's function for this problem is the function -.i i-kvi,-;

the equation


2 2 a2 a2 2
A2 X) ( + )( + = 6( ')
dx2 ay2 dO 2 ay2

where 6( 7, 7') is the Dirac delta function and 7 = (x, y) and 7''

This can be solved with the solution


U(r) =r2logr2,


(2.5)

(x', y') E R2.


(2.6)


where r is the distance 2 +y2.

Because an affine transformation has a zero contribution to the bending

energy, the transformation allows a free affine transformation. Define the matrices


0

U(r21)



U(rl)


U(r12)

0



U(rr2)


.. U(rin)

U(r2.)


0


, which is n x n;


, which is 3 x n;


(2.7)


(2.8)


and
K P
L which is (n + 3) x (n + 3), (2.9)
PT 0

where the symbol T is the matrix transpose operator and O is a 3 x 3 matrix of

zeros.









Let V = (vi,..., v,) be any n-vector and write Y = (V 1 00 0). Define the

vector W = (w, ..., ,). and the coefficients (ai, a,, ay) by the equation


L- Y =(W la a, a,)T. (2.10)

Finally we have the solution


f(x, y) = a + axx + ayy + (|P,- (x,y)|). (2.11)
i=1

2.3 More on Splines

Bookstein's seminal work applying thin-plate splines to the landmark problems

in 2D [ ] is well known in the computer vision and imaging community. What is

less known is the theoretical foundations for the thin-plate spline laid by Duchon

[ ]and Meinguet [ ]in the 1970s.
This solution is correct and simple. However it is not numerically stable

because it involves the inverse of the large kernel matrices. A numerically stable

solution is given by Wahba [ ] using the QR decomposition. Wahba has an

account for a more generalized formulation for d-dimensional space and the energy

involving mnh order partial derivatives. Here we only discuss the special thin plate

spline with d = 2 and m = 2. Later in our work in 3D situations, 3D thin plate

splines are used where d = 3 but it is straightforward to generalize from 2D to 3D,

while the kernel in 3D is different than that for 2D.

Let f, g E Wk'2(Q), where Wk'2(Q) is the Sobolev space. The inner product of

f and g is defined as

A thin-plate smoothing problem in 2D is to find f E Wk'2(Q) to minimize the

functional



il- f(P 2 + AE(f). (2.12)
i= 1









The null space is spanned by



1(t) = 1, 2(t) 3 (t)= y,


which is an affine space.

Duchon [ ] proved that if {pi E Qli i

regression on 1, 2, 3 is unique, then 2.12

representation


S1, 2, ..., n} are such that least squares

has a unique minimizer f, with the


f(t) M= d,,(t) + ci K(t, pi), (2.14)
v=l i =
where K is a Green's function for the double iterated Laplacian. K has the form


K = K(Is tl),


K(r) = r2og T| .


By using integration by part, we obtain that c, d are the minimizers of


- y Td- Kc 2 + Ac'Kc
n


subject to T'c = 0, where K is a n x n matrix with ijth entries K(pi, pj). Let us do

a QR decomposition on T


R
T (Q1 : Q2) (2.18)
0

where (Qi, Q2) is orthogonal and R is lower triangular. Qi is n x 3 and Q2 is

n x (n 3). Since T'c = 0, c must be in the column space of Q2, with c = Q27 for

some n 3 vector 7. Substituting in 2.17 and the energy is in the form of


(2.13)


and


(2.15)


(2.16)


(2.17)











Q'2y Q Q 12 + Q'y -Rd Q 2KQ2- +2 X 'Q KQ2. (2.19)
n n

The solutions are

d = R- (y KQ2)

c = Q2, (2.20)

with

S=(Q2KQ2 +nA 1)-lQ2 (2.21)

where I is the identity matrix.
2.4 Reproducing Kernel Hilbert Space (RKHS) Formulation

It is possible to come to the same result for the minimizer using reproducing

kernels and from this point of view it is easier to understand the existence and
uniqueness of the solution [ ].

Let I| lf2 = E = f(fl + 2f2~ + f y,'/. where Ilfll is the norm of f in
Wk,2(Q). Since Wk'2(Q) is a Hilbert space, from Riesz representation theorem, for

any x E Q, the evaluation linear functional



S: Wk2(Q) R,

(f) = f() (2.22)

has a representer ux E Wk,2(Q) such that


f() =< u, f >. (2.23)

Now the problems is transformed to a new problem: find a function f E
Wk,2(Q) with minimal norm I|/f subject to constraints











< Ux f > vi, i 1, 2,...,n. (2.24)

For Xa, Xb E Q, u(Xa, Xb) UX, (Xb) is the kernel of the reproducing kernel

Hilbert space.

Let T be the linear subspace spanned by u,, i 1, 2, ..., n. Any function

f e W',2(Q) can be decomposed into f = fT + fi where fT E T and fi is in the

orthogonal complement of T and hence < ux,, fi >= 0. We know if fr satisfies

(2.24), then f also satisfies (2.24) only with | f|| > ||fTr| if fi 0. So we only need

to search for the solution in T. The general solution thus can be written as

n
f(X) =ao + aix + a2y + ,(X,, X),
i=1
where functions of the form ao + aix + a2y span the null space. With this form, E

can be rewritten as



SUijj = WUW+. (2.25)
i=1,j=1

2.5 The Folding Problem of Splines

This is all very nice except as mentioned above there is no mechanism to

guarantee a bijection in order to be homeomorphic or diffeomorphic. Intuitively

this problem is known as the folding of space.

Figure 2-2 shows the displacement of landmarks. Figure 2-3 is the thin-plate

spline interpolation. We can see the folding of space. This is the drawback of thin-

plate spline interpolation. Due to the folding of space, features in the template may

be smeared in the overlapping regions. And furthermore, the transformation is not

invertible. A diffeomorphic transformation is strongly desirable, since it preserves

the features, the topology and is smooth as shown in Figure 2-4.









2.6 Imposing Restriction on the Jacobian

One straightforward approach to find a diffeomorphism for practical use is to

make some remedy on the thin-plate spline. We can restrict our search space to the

set of diffeomorphisms and the ideal one should minimize the thin-plate energy. We

make the observation that if the Jacobian of the transformation f changes sign at a

point, then there is folding. We put the constraint requiring the Jacobian is always

positive. There is some literature on this approach but most of these approaches

does not guarantee that the transformation is smooth everywhere [ ].

2.7 The Flow Approach

Another approach is to utilize a flow field[ ]. We introduce one param-

eter, the time t into the diffeomorphism. Let Ot : f -Q f be the diffeomorphism

from Q to f at time t. A point x is mapped to the point Qt(x). Sometimes we

also denote this as O(x, t). It is easy to verify that for all the values of t, Ot form

a one parameter diffeomorphism group. If x is fixed, then O(x, t) traces a smooth

trajectory in 2. The interpolation problem becomes: find the one parameter diffeo-

morphic group 0(-, t) : 2 -- 2 such that given pi E 2 and qi E 2 Vi = 1, 2, ..., n,

q(x, 0) = x and q(pi, 1) = qi. We introduce the velocity field v(x, t) and construct a
dynamical system using the transport equation

0((x, t), ). (2.26)
ot

The integral form of the relation between O(x, t) and v(x, t) is



q(x, 1) + j v( (x,t),t)dt. (2.27)

Obviously, such a q(x, t) is not unique and there are infinitely many such solutions.

With the analogy to the thin-plate spline, it is natural that we require the desirable

diffeomorphism results in minimal space deformation. Namely we require the

deformation energy










Si Lv(x, t)2dxdt (

to be minimized, where L is a given linear differential operator.
The following theorem [ ]states the existence of such a velocity field and
shows a way to solve for it.
Theorem 4. Let pi E Q and qi E Q Vi = 1, 2,..., n. The solution to the ,.

minimization problem


v(.) arg min I Lv(x, t) 2dxdt
J1onJ


subject to


Vi 1, 2, ..., n


(2.30)


x+ I v( (x, t), t)dt (2.31)

(.-, 1) : Q --+ The optimum .... /:11 field i
by


v(x, t)= K((Xi(t), x) Z (K(( (t))-) ,(x, t)
i j=1




SK(p(pi, t), (pi, t)) K(O(pl, t), (pn,t


(2.32)


K(O(t))


2.28)


i,'/


(2.29)


where


p(pi, 1) q- i,


1(x, 1)

exists and .7. ;i:. a diffeomorphism
and the diffeomorphism ( are given


where


(2.33)


SK(O(pn, t), (pi, t))


K(O(p., t), (p., t)) /









with (K((q(t))ij denoting the ij, 3 x 3 block entry (K(((t))ij = K(q(pi, t), c(pj, t)),
and



(pn, ) = arg min) j (p, t)T(K((t))-1 .)(pj,t)dt (2.34)
ij

subject to ((pi, 1) = q i = 1, 2, ..., N with the optimal diffeomorphism given by


q(x, 1) x+ v((x,t)t)dt. (2.35)

The proof [ ]is omitted here. With this theorem, we can convert the original

optimization problem looking for the vector field v(x, t) to a problem of finite

dimensional optimal control with end point conditions.

This problem is called the exact matching problem because we required the

images of the given points pi, i = 1, 2, ..., n are exactly another set of given points

qi, i 1, 2, ..., n. The exact matching problem is symmetric with respect to two
sets of landmarks or two point shapes. Swapping the two point sets {pi E Q l| i

1, 2, ..., n} and {qi E G 21 i 1, 2, ..., n}, results in the new optimal diffeomorphism

to be the inverse of the old diffeomorphism.

The exact matching problem can be generalized to the inexact matching

problem. In the inexact matching problem, we do not require the points exactly

match. Instead, we seek a compromise between the closeness of the matching points

and the deformation of space. We minimize


SlilL(x, t) l2dxdt +A lqi- o(, 1) 2, (2.36)

which can be similarly solved.

As seen from the formulation of the problem, there may be infinitely many

diffeomorphisms that interpolate the two sets of landmarks. A usual way to find a

particular desirable diffeomorphism is to require the diffeomorphism to minimize a









certain objective function. Camion and Younes [ ]proposed a different objective

function in the form of



E(v, q)= Lv(x, t)2 dxdt+ dt v(q(t)) t (2.37)
i 1

over all time dependent velocities v(x, t) on C and over all trajectories qg(t),..., q,(t).

This can be interpreted as a geodesic distance between two points on the configura-

tion manifold. This is generalized to exact matching by Marsland et al. [ ].

2.8 Correspondence and Softassign

So far all work has assumed landmark matching with a known correspondence.

When the correspondence is unknown, the problem is dramatically complicated.

One approach is the softassign [ ] method, which solves for the correspondence

as a permutation problem using linear programming in the space of doubly

stochastic matrices [ ]. The problem of finding the correspondence between two

sets of points can be formulated as finding the permutation matrix.

This approach is expensive in computational time. Chui et al. [ ,

] adopted the joint clustering scheme, in which the correspondence and space

deformation is estimated simultaneously. However, in all the work, splines are

used and a potential drawback is that a diffeomorphic mapping in space is not

guaranteed.

2.9 Distance Transforms

An unusual way of getting around the point correspondence problem is to use

distance transforms to convert the point matching problem to an image matching

problem. The distance transform was first introduced by Rosenfeld and Pfaltz

[ ], and it has a wide range of applications in image processing, robotics, pattern
recognition and pattern matching. Paragios et al. [ ] give one example of using

distance transforms to establish local correspondences for compact representations









of anatomical structures. Distance transform applies to binary images, as well as

point sets, which can be thought as a special case of a binary image. Suppose we

have a domain 2, and a point set S C 2. For each point x E 2, we assign to it a

non-negative real number, which is the shortest distance from x to all the points

in S. This way we obtain a scalar field in domain 2. We treat this scalar field as

a gray scale image, we call it the distance transformed image of the point set. If

we have two point sets, we can first perform the distance transform on the two sets

respectively and later register the two distance transformed images. However, there

are problems with this approach. If the matching of the two distance transformed

images is intensity based, then the original points may not be matched exactly.

Even in the inexact matching case, the optimization for the distance transformed

images is over the entire image region and that may not be optimal for the original

point set. If the matching of the two distance transformed images is level set based,

the level sets in the two images may not be topologically equivalent, as shown

in Figure 2-5 and Figure 2-6. In Figure 2-5, there are two distance transformed

images, with three points each in the image. Figure 2-6 shows the level sets of the

two distance transformed images. Hence this method cannot guarantee to obtain a

diffeomorphism.

Due to the indirect approach of transforming point sets into distance trans-

forms, this method has not seen wide applicability for point sets.

2.10 Implicit Correspondence

It is possible to define distance measures between two shapes, with the shapes

viewed as point sets, without the knowledge of correspondence. Hausdorff distance

between two point sets is a well-known distance measure for such purposes.

Huttenlocher et al. [ ] develop methods comparing shapes using Hausdorff

distance. Glaunes et al. [ ] use another method to circumvent the correspondence

problem by introducing a distance measure of the two shapes using an external









Hilbert space structure. If the two shapes are point sets, the distance measure

between two point sets used in Glaunes et al. [ ]is

1
29|4> v|. (2.38)
UR

where [|p2| is the norm squared in some Hilbert space I* and

N1 N2
\=k cj2ki(x*,yj) (2.39)
i=l j=1

where ci, cj are constant coefficient. xi are the points in the first sets and yj are the

points in the second sets. ki is some kernel. In their experiments, they used radial

basis function kernels


xI y12
ki(x,y) = f( ), (2.40)

with fi(u) = e- and fi(u) = 1

The idea of distance measure without explicit correspondence actually goes

back to Grimson et al. [ ]and Lu and Mjolsness [ ]in 1994. The distance

measure used is
ITli-m |2
EE C 2,2 (2.41)
i j
As pointed out in Riingirajiian et al. [ ]this is equivalent to using



E(M, T) = a 3 Di (T) + 3 log 1 (2.42)
ij ij
where Mij is the explicit correspondence and T is the space transformation and Dij

is the distance measure between point pairs that are in correspondence. Guo et al.

;-ii.-.- -- I1 a joint clustering algorithm for solving 2D diffeomorphic point matching

problems, with unknown correspondence, using explicit correspondences [ ].

The different formulations with explicit correspondence and implicit correspondence









are closely related to each other, through a Legendre transform, as pointed out by

Mjolsness and Garrett [ ].

2.11 Shape Context

In the approach of shape context [ ], the correspondence is expressed explicitly

and the correspondence problems is tackled directly. Moreover, the correspondence

is solved separately from the space transformation with the help of shape context.

Shape context is the local shape information at each point. For each point q and all

other points pi, we can draw a vector from q to pi, that is pi q. The local shape

context information is all stored in this set of n 1 vectors. This information is

rich and in practice the distribution of these n 1 vectors gives us more robust,

compact and discriminative descriptor. For each point pi, we compute a histogram

hi of these n 1 vectors


h (k) #{q / pi (q p) E bin(k)}. (2.43)

The histogram is defined to be the shape context of point pi. Bins that are uniform

in log-polar space are used to make the nearby context points more important the

the far away context points. The cost for the correspondence between two shapes is

the sum of the cost of corresponding point pairs


H() = C(pi, q,(i)), (2.44)

which is a function of the permutation r, and the cost with each individual pair of

points is defined as


Cj -C(p, qj) h(- (2.45)
C2 k1 i(k) + hj(k)


2.12 Active Shape Models

Cootes et al. propose a method, which they call Active Shape Models [

to locate objects in images, with the help of model shapes and the training of









these model shapes. This approach plays as a bridge between the landmark based

methods and image intensity based methods, and also a bridge between rigid

matching and non-rigid matching. They use landmark points on the models but

not on the test images. The landmarks are hand picked and hand labeled. The

automatic correspondence problem is circumvented by the human involvement of

the landmark picking and labeling process. The deformation model of the template

is similarity, which is very close to rigid, allowing translation, rotation plus a

scaling. What makes it applicable to non-rigid deformation is that they learn the

statistics of the model shapes through a set of training samples and find the mean

and variance in higher dimensional space and apply the PCA analysis with the

variances. When it applies to locating the model in the image, they use the snake

model, which is non-rigid in nature, with the help and restriction of the statistics of

the model shapes.

2.13 Deterministic Annealing Applied to EM Clustering

Deterministic annealing is an effective technique used in the clustering

problems. The clustering problem is a non-convex optimization problem. The

traditional clustering techniques use descent based algorithms and they tend to get

trapped in a local minimum. Rose et al. [ proposed an annealing approach using

analogies to statistical p1 l-i. -

The clustering problem is to partition a set of data points, {xi E Rd i

1, 2,..., n} into K clusters C1, C2,..., CK, with the centers rl, r2,..., TK respectively.

In the fuzzy clustering literature, we call the probability of each point pi belonging

to each cluster Cj the fuzzy membership. Hard clustering is a marginal special

case, where each point is deterministically associated with a single cluster. Let

Ej(xi) denote the energy associated with assigning a data point Xi to cluster Cj.









The average total energy is


< E >= P(x G Cj)Ej(xi). (2.46)
i J

Since we do not make any assumption about the data distribution, we apply

the principle of maximum entropy. It is well known from statistical ]lli'--i. -. the

association probabilities P(x, C Ci) that maximize the entropy under constraint

(2.46) are Gibbs canonical distributions,


P(x CE C) ,,e- (2.47)
Zi

where Zi is the partition function


Z, E (2.48)
k
The parameter / is the Lagrange multiplier determined by the given value of

< E > in (2.46). In the analogy in statistical plr, -i. -. 3 is inversely proportional

to the temperature T. We have assumed that we have a fixed set of clusters. We

want to extend this to include optimization over the number of clusters as well.

The optimal solution then is the one that minimizes the free energy

1
F In Z. (2.49)


The set of cluster centers are the ones that satisfy

OF
So0, Vj. (2.50)
Orj

And the solution is

r E=( ) (2.51)
Y3 EPXi P C, Cj)
This procedure determines a set of cluster centers {rj E Rd j = 1, 2, ..., n} for

each fixed 3. Generally, changing K, the imposed number of clusters, will modify

the positions of the set of cluster centers. However, there exists some n, such that









for all K > rn, one gets only n, distinct cluster centers while the remaining K n,

cluster centers are repetitions from this set. Thus at each given 3 we get at most rn

clusters. Here we assume K > n, at a given 4 and we only consider them without

repetitions.

The free energy F and 3 are Legendre transform images of each other. Fixing

one of them determines the other. For / = 0, each data point is uniformly

associated with all clusters and all the centers have the same location, the centroid

of the data. Clearly, for 3 = 0 we have a single minimum, which is the global

minimum, for F, and the entire data set in interpreted as one cluster. At higher 3,

the free energy may have many local minima, and the concept of annealing emerges

here can be viewed as tracking the global minimum while gradually increasing P.

Moreover, at 3 = 0 there is only one cluster (rn = 1), but at some positive 3 we

shall have n, > 1. In other words, this cluster will split into smaller clusters, and

will thus undergo a phase transition. The first phase transition occurs at a critical

value for /

Pc = -(2.52)
2max

where Amax is the largest eigenvalue of Cxx, the covariance matrix of the data

set. Given = 1/T and for Gaussian mixture clustering T = 2
understandable that the critical value a( for the prescribed or is a c = /max.

2.14 Statistical Shape Analysis on Differentiable Manifolds

David Kendall first introduced the idea of representing shapes in complex pro-

jective spaces. The idea is developed to represent shapes on general differentiable

manifolds, or shape spaces [ ]. A shape is constructed from a sequence of land-

marks pi, l_, ...p, with pi E R2. Namely each landmark is a point in 2-dimensional

Euclidean space. The two sequences of landmarks in R2 are considered of the same

shape if they differ only by a similarity transformation in R2. A shape then is









defined as an equivalent class of these landmark sequences and the shape space is

carved out of the quotient space defined by the equivalent relation.

One way to distill the shape information out of the landmark sequence is to

remove the location, scale and orientation. To remove the location, we can define

r, = x (2.53)

where

x (2.54)
j=1
is the mean or centroid of the landmarks. That is, to remove the location, we make

the landmarks mean zero. To remove the scale we can make the variance of the

landmarks as one. So we define


T x- (2.55)
1:J21 IIXj- X112

We refer the vector T as the pre-shape of the landmarks. The pre-shape space is the

intersection of the (n 2)-dimensional subspace

F2n-2 {(*X, ...X,) R2n | Y x = 0} (2.56)
j=1

with the unit sphere

S2n-1 {(X, ..., n) R2n 1, 2 = 1}. (2.57)
j=1

The intersection

S2n-3 2 F2n-2 S2n-1 (2.58)

is a (2n 3)-dimensional sphere within the ambient Euclidean space R2".

It is more difficult to remove the orientation of the shape. To do this, we

define the orbit of a pre-shape Tr S2'-3 is the circle


0(r) = {0(r) 1|0 < 0 < 2r} C S2"-3


(2.59)









Two pre-shapes are of the same shape if they are on the same orbit. A shape

is defined as the equivalent class of the pre-shapes. If Tr and T2 are two preshapes,

then the great circle distance between T1 and T2 is given by


d(TI, T) = COS-' T1, 7T >). (2.60)

The induced metric on ZE is then defined as


d[O(Tr), O(T2)] infd([1(71), 02(T2)] 0 < 01,02 < 27. (2.61)


This is called Procrustean distance. It can be proved that if ri and T2 are two

representatives of two shapes, the procrustean distance between the two shapes can

be expressed as

d(ai,a 2) COSi( T71k2k). (2.62)
k=1
The standard statistics like means and variances can be performed on the

shape manifold.

One drawback of this approach is that the shape is treated as the sequence of

landmarks instead of sets. That means if we two identical sets of landmarks and

only label them differently, this theory treats them as two distinct shape. It does

not consider to make an equivalent class from the permutation of landmark points.

2.15 Distance Measures from Information Theory

Information geometry is an emerging discipline that studies the probability

and information by way of differential geometry. In information geometry, ev-

ery probability distribution is a point in some space. A family of distributions

corresponds to points on a differentiable manifold.

Endres et al. [ ] proposed a metric for two distributions. Given two proba-

bility distributions P and Q, and R = (P + Q), we can define a distance DpQ









between the two distributions by

DQ = 2H(R)- H(P)- H(Q)

SD(P R) + D(QI R)
Ni- (plog 2pi 2pi
= (plog + log ) (2.63)
i1 Pi qi Pi + qi

This metric can also be interpreted as the square root of an entropy approximation
to the logarithm of an evidence ratio when testing if two samples have been drawn
from the same underlying distribution. 1DpQ is named Jensen-Shannon divergence,
which is defined as

D,(P,Q) = AD(PIIR)+ ( )D(QIIR) (2.64)

R = AP (1 A)Q

and therefore

-D, D(P,Q). (2.65)
2 Q D
With probability distribution P(x|0) where 0 is a set of parameters 01,..., On,
the Fisher information is defined as [ ]

2 alogp(x 8) a2 logp(x 8)
Gij(0) = -E [0 () 0= l pgP ) dx. (2.66)
00 ,[Oj -O 01,Oj

C. R. Rao [ ] -ii.-. -led this is a metric. In fact, it is the only suitable metric in

parametric statistics and it is called Fisher-Rao metric.


































0.4


0.2


01


-0.2



20

15

10
5
5


Figure 2-1: Deformation of the thin plate







100

90

80








30
70








20-

10

0
0 20 40 60 80 100


Figure 2-2: Landmark displacements































0/
80-
8i .

60 -


40 -.


20-


0 20 40 60 80 100 120




Figure 2-3: Thin-plate Spline interpolation


0 20 40 60 80 100 120 140




Figure 2-4: Diffeomorphic interpolation

























300, -W



200



100




0 100 200 300 400 500 600

(a)
600



500



400




300-



200




100




0 100 200 300 400 500 600

(b)

Figure 2-5: Two distance transformed images of three landmarks





































0 100 200 300 400 500 600
(a)
300
















400 -
200










00




0 100 200 300 400 500 600
(b)

Figure 2-6: Level sets of two distance transformed images















CHAPTER 3
DIFFEOMORPHIC POINT MATCHING

In this chapter we investigate the diffeomorphic point matching theory and

apply the theory to shapes in medical imaging. In Section 3.1 we prove a theorem

about the existence of a diffeomorphic mapping matching the landmarks. In

Section 3.2, we prove another theorem about the symmetric nature in the case of

exact landmark matching. In Section 3.3 we formulate a theory of diffeomorphic

point matching with unknown correspondence and devise an objective function.

In Section 3.4 we design an algorithm to solve the problem using joint clustering

and deterministic annealing. In Section 3.5 we apply the algorithm to 2D corpus

callosum shapes. In Section 3.6 perform matching on 3D hippocampus shapes.

Both corpus callosum and hippocampus are parts in the human brain and the

matching of these shape have great significance in medical treatment and medical

research.

3.1 Existence of a Diffeomorphic Mapping in Landmark Matching

We have discussed in Chapter 1 that in 2D a similarity transformation

can map exactly 2 given points to 2 given points. A affine transformation can

map exactly 3 given points to 3 given points. A projective transformation can

map exactly 4 given points to 4 given points. Now given n arbitrary distinct

points {pi E R2 i = 1, 2, ...n} and another set of n arbitrary distinct points

{qi E R21 = 1, 2, ...n}, we want to find a diffeomorphism f : R2 -- R2 such

that f(pi) = qi. It is natural to ask the question, does such a diffeomorphic

mapping always exist? Our intuition is it exist and there are infinitely many such

diffeomorphisms. This is stated as our first theorem and the proof follows.









Theorem 5. A diffeomorphic transformation that interpolates arbitrary number of

n pairs of landmarks always exists.

Proof We show the existence by construction. We construct a simple, although

most likely undesirable in most of the applications, diffeomorphism. The intuitive

idea is to "dig canals" connecting the landmark pairs. We first choose the first pair

of landmarks pl and qi. Assume no other landmarks lie on the line connecting pl

and ql. Establish a coordinate system such that pi and ql are on the x axis. Let

the signed distance from pi to ql be d. Construct the transformation fl : 1 --- 2

such that fi(x, y) = (x', y'),


= + de-"2 (3.1)

y' = y

where v = tan(-y), for any arbitrarily small e. It is easy to show that fl is

a diffeomorphism and it maps pi to ql and keeps all other landmarks q2,..., q,

fixed. This is very much like the flow of viscous fluid in a tube. Similarly we can

construct diffeomorphism f, that maps pi to qi and keeps all other landmarks fixed,

for i = 1, 2, ..., n. The composition of this series of diffeomorphisms


f f O ... f2 fl (3.2)

is also a diffeomorphism and obviously f maps pi to qi, for i = 1, 2,..., n.

If some landmark qk lies on the line of pi and qi, we can find such a direction

such that we draw a line lk through qk and there are no other landmarks on the

line. Then we make a diffeomorphism h transporting qk to a nearby point q, along

the line without moving any other landmarks, using the same canal of viscous

fluid technique. Then we make diffeomorphism fi as described before. After that,

we move landmark q' back to the old position with the inverse of h-1. So we use

Fi = h- fih in place of fi.






















_0


Figure 3-1: Existence of a diffeomorphic mapping




3.2 Symmetric Matching due to Time Reversibility

Asymmetry exist in many image and shape matching situations. Suppose we have a

point set {pi E R2 i = 1, 2,...n} and another point set {qi E R2 i = 1, 2,...n}, we

find the diffeomorphic mapping f : R2 -- R2 which minimize the energy functional

E(f) subject to the constraints f(pi) = qi. Now we define a reverse problem,

namely to find a mapping g : R2 R2 which minimize the same energy functional

E(f) subject to the constraints f(pi) = qi as shown in Figure 3-2. In general,

depending on the objective function to minimize, g / f-1. This is called

..i-. mmetry of the matching. In some cases we do have g = f-1. The matching is

called symmetric then and this is a nice property to have. The following theorem

states that for the exact diffeomorphic landmark matching case, the matching is

symmetric due to the time reversibility of the flow.









f



S g g **


g

Figure 3-2: Asymmetry of the matching

Theorem 6. If (xk, 1) y k and O(x, t) and v(x, t) minimize the i ',. it

E= Lv(x,t) 2dxdt,

then the inverse "'":I*',':./ maps the landmarks backward -l1(yk, 1) Xk and
0-l(x, t) and -v(x, -t) also minimize the energy E.

Proof. First, from the known property of the diffeomorphism group of such a
dynamical system, O(x, tl + t2) = 2((x, t) t), it is easy to show that -1 (x, t)

(x, -t). This is because (., -t) o (., t)(x) = (., t) o (., -t)(x) = (((x, t), -t)
(x, t + (-t)) = (x, 0) = x. And O(x, -t)and -v(x, -t) also satisfy the transport
equation = -vv(o(x, -t), -t). Suppose O(x, t) and v(x, t) minimize the
energy E = f f ILv(x,t) |2dxdt, but -'l(x, t) (x, -t) and -v(x, -t) do not
minimize the energy E =fo f |Lv(x, t) 2dxdt. Let the minimizer be y(x, t) and
u(x, t) such that Vk, y(yk) Xk and o fd IILu(x, t) 2dxdt f f ILv(x, t)2dxdt.
Then, we can construct (x, t) = (x, -t) such that (x, t) and -u(x, -t)
satisfy the transport equation and -'(ZXk, 1) Yk. However fo1 f I ILu(x, t) 2ddt

j J~ f IILv(x, t)2dxdt contradicts the assumption that v(x, t) is the minimizer of
the energy E.









3.3 A Theoretical Framework for Diffeomorphic Point Matching

There are different ways to solve for the unknown point correspondences [ ,

]. Essentially, within the framework of explicit point correspondences-as
opposed to the distance function framework of implicit correspondence-we have a

choice between i) solving for an optimal permutation and ii) letting corresponding

"labeled" points discover their optimal locations. We opt for the latter in this work

because of its simplicity. The clustering in fact serves two purposes. First, it is the

method to find the unknown correspondence. We initialize the two sets of cluster

centers around the centroids of their data points, respectively. The cluster centers

are labeled with identical labels in the two sets denoting correspondence. The

cluster centers evolve during the iterations of an incremental EM algorithm and

they are linked by a diffeomorphism and are forced to move in lock-step with one

another. Second, clustering is the modeling of the real data sets, with noise and/or

outliers because with two shapes represented by point samples, we cannot assume a

point-wise correspondence. The correspondence is only between the two shapes and

clustering is a useful way to model the shapes.

We use a Gaussian mixture model to describe the clustering of the point

sets. For more details on this approach along with justifications for the use of this

model, please see [ ]. The Gaussian mixture probability density is


1N I1 1
p(x|r,a) a= (2iare)d/2 exp( 2- |X Tk 2) (3.3)

with x being a point in Rd, r as the collective notation of a set of cluster centers

and oa as the variance of each Gaussian distribution. The reason for the notation

of subscript T will be discussed in Section 3.4 in the context of annealing (T will

be the temperature. aT and T are related by T = 2o.) Here we just understand

a( as the prescribed variance in the Gaussian mixture model as opposed to the

actual measured variance a2 from data fitting.










The clustering process is the estimation of the parameters r that leads to the

maximum log-likelihood of the observed sample

N1 N
logp(xlr, aT) log exp(-- || i rI 2). (3.4)
i=i k=1

The solution can be found by applying the EM algorithm. As pointed out by

H.il h.,:.iv [ ], in the mixture model context, the EM algorithm maximizing (3.4)

can be viewed as an alternative maximization of the following objective

Ni N
F(M, r) = !, x- rk2 (3.5)

Ni N
Y logic ,.
i=-i k=1

This is equivalent to minimizing


E(M, r) -F(M,r) (3.6)
SNi N

T i=1 k=1
Ni N
+ Y ; log I
i=- k=1

with simplex constraints on M.

The clustering of the other point set is identical. For the joint clustering

and diffeomorphism estimation, we put together the clustering energy of the two

point sets and the diffeomorphic deformation energy induced in space giving us an









objective function

E(/',/ ,r,s, v,O) (3.7)
1 N1 N N1 N
Y YE ; Xi 2 I E ;'/, log',
i=-1 k 1 i=l k=1
N2 N N2 N
M1 k+y Y Sk2 ifOg Mjk
jT l k=l j=1 k=1

+ -7', 1 1)1 A' I \\Lv(x, t12dxdt.
k 1

In the above objective function, the cluster membership matrices satisfy

j\, E [0,1], 1Vik MJYk [0, 1], Vjk and E7 ', 1, E 1 Mfk 1. The
matrix entry /', is the membership of data point xi in cluster k whose center is at

location rk. The matrix entry Mfk is the membership of data point yj in cluster k

whose center is at position Sk.

The diffeomorphic deformation energy in f is induced by the landmark dis-

placements from r to s, where x E f and O(x, t) is a one parameter diffeomorphism:
Q -- Q. Since the original point sets differ in point count and are unlabeled, we

cannot immediately use the diffeomorphism objective functions as in Joshi and

Miller [ ] or Camion and Younes [ ] respectively. Instead, the two point sets

are clustered and the landmark diffeomorphism objective is used between two

sets of cluster centers r and s whose indices are always in correspondence. The

diffeomorphism q(x, t) is generated by the flow v(x, t). q(x, t) and v(x, t) together

satisfy the transport equation "= v( (x, t), t) and the initial condition Vx,

O(x, 0) = x holds. This is in the inexact matching form and the displacement
N 12
term S1 I Isk (rk, 1)2 plays an important role here as the bridge between the

two systems. This is also the reason why we prefer the deformation energy in this

form because the coupling of the two sets of clusters appear naturally through the

inexact matching term and we don't have to introduce external coupling terms as









in Guo et al. [ ]. Another advantage of this approach is that in this dynamic sys-

tem described by the diffeomorphic group q(x, t), the landmarks trace a trajectory

exactly on the flow lines dictated by the field v(x, t). Also, the feedback coupling

is no longer needed as in the previous approach because with this deformation

energy described above, if q(x, t) is the minimizer of this energy, then (x, t) is

the backward mapping which also minimizes the same energy.

The aT in (3.7) is a fixed parameter. It is not a variable during the min-

imization. It is an attribute of the point set and can be a priori estimated for

the point shapes. The reason that we have coefficient -7- in front of the term

Ek1 1 8k (rk, 1)112, instead of another free parameter is that oT is a natural unit
of measurement for distant discrepancies in clustering and it does not make sense

to make this coefficient too big or too small. Since aT is a constant, we multiply

the objective function by a constant 2UT and we get the final form of the objective

function as



E (\ ',\! ,r,s, v, ) (3.8)
N1 N Ni N
u i; -, Tk 2 + 2,7T2UY /;< VlogJ V,
i= k=l1 i=l k=1
N N2 N
ZZMT2Yk iZk log sM Y
M EjkJii- Sk +2 2 k10g k
k=l j=1 k=1
N 1
+ YIS (rk, 1)11 +2 A ] v t) 2 dxdt.
k=1 T 0 1

3.4 A Diffeomorphic Point Matching Algorithm

Our joint clustering and diffeomorphism estimation algorithm has two compo-

nents: i) clustering and ii) diffeomorphism estimation.

For the clustering part, we use the deterministic annealing approach. The

clustering problem is a non-convex optimization problem. The traditional clustering









techniques use descent based algorithms and they tend to get trapped in a local

minimum. Rose et al. [ ] proposed an annealing approach using analogies to

statistical p lli-i, The clustering cost function is seen as the free energy of a Gibbs

canonical distribution. The minimization of clustering cost function is seen as

the simulation of a 1'li. -i 1 .1 annealing process in which free energy is minimized.

Let T be the temperature of the system and in the clustering system T = 2o4.

Let 3 = 1/T be the reciprocal temperature. Initially let 3 = 0. We have a

single minimum for the energy in (3.6), which is the global minimum and all the

cluster centers are located at the same point, which is the center of mass of all the

data points and each data point is uniformly associated with all clusters. In the

numerical implementation, we initialize all the cluster centers on a sphere of a very

small radius and there are no data point within the sphere. When the temperature

is lowered gradually, at a certain critical value of temperature, the clusters will split

into smaller clusters and a phase transition occurs. At lower T, the free energy

may have many local minima but the annealing process is able to track the global

minimum.

For the diffeomorphism estimation, we expand the flow field in term of the

kernel K of the L operator

N
v(x, t) = -ak(t)K(x, Ok(t)) (3.9)
k=1
where Ok(t) is notational shorthand for ((rk, t) and we also take into consideration

the affine part of the mapping (not written out in the above equation) when we

use the thin-plate kernel with matrix entry Kij = ri log rij for 2D and Kij = -rij

for 3D, with rij =II zi xj II. After discretizing in time t, the objective in (3.7) is

expressed as













N1 N

i=-i k=1
N2 N

j=1 k=1


N1 N
rkI1 + T Y JM ', logMJ ,

N2 N
Sl k k lg Mlk
j=1k=1


N N S
+ E I- rk [P(t)(t) + a,(t)K(k (t), 21(t))] I12

N N S
S+ STS- < k(t),al(t) > K(Ok(t),0l(t))
k=1 1=1 tO0


where


1 v(t) 2 (t) 3 (t)


1 <1(t) <2(t) 03(t)


(3.11)


and d is the affine parameter matrix. After we perform a QR decomposition on P,


R(t)

0 )


(3.12)


We iteratively solve for ak(t) and Ok(t) using an alternating algorithm. When Ok(t)
is held fixed, we solve for cOk(t). The solutions are


Qi(t)K(O(t))Q2(t)7(t)]


(3.13)


a(t) = Q2(t)7(t)


(3.10)


P(t)


E( 1', 1 ,r,s, a(t),0(t))


d(t) = R-l(t) [Ql(t)O(t + 1)


(3.14)


P(t)=- (Q(t) : Q2(t)) (








where K(O(t)) denotes the thin-plate kernel matrix evaluated at

0(t) {f (k,t)Ik 1,...,N} and

7(t) (Q (t)K((t))Q2(t) + T)Q( (t 1). (3.15)

When ak(t) is held fixed, we use gradient descent to solve for Ok(t):

OE N
O* = 2 < ak (t), c(t) 21i4 > V I(Ok(t), 1(t)) (3.16)
c.' (t) l=1
where = s -_ rl E o am(t)K(0,mn(t), 0p1(t))dt.
The clustering of the two point sets is handled by a deterministic annealing
EM algorithm which iteratively estimates the cluster memberships Mx and My and
the cluster centers r and s. The update of the memberships is the very standard
E-step of the EM algorithm [ ]and is performed as shown below.


exp(- ,|| I k 112)
/ Vi-----k vi(3.17)
1l1 exp(-- || -- ri|2

I7 exp(- l/|| s, I-)
Mfk N Vjk.. (3.18)
E Iiexp(- .,||,/ /-s1||2)
The cluster center update is the M-step of the EM algorithm. This step is not the
typical M-step. We use a closed-form solution for the cluster centers which is an
approximation. From the clustering standpoint, we assume that the change in the
diffeomorphism at each iteration is sufficiently small so that it can be neglected.
After making this approximation, we get


S1 'k Sk EI 1 o al(t)K(l((t), Ok(t))dt
1 C (3.19)

Y1 Mkyj + (rk, 1)
j 1k k V. (3.20)
3- Mik1 j


The overall algorithm is described below.









Initialization: Initial temperature

T = 0.5(1ii.v: ;l ,. I|- + maxj Il; / IJ-) where x, and y, are the centroids of

X and Y respectively.

Begin A: While T > Tfinal

Step 1: Clustering

Update memberships according to (3.17), (3.18).

Update cluster centers according to (3.19), (3.20).

Step 2: Diffeomorphism

Update (q, v) by minimizing


C
Ediff(, v) = Sk-(rk, 1)12
k=l

+ AT IILv(x,t) 2dxdt


according to (3.13)(3.14) and (3.16).

Step 3: Annealing. T 7T where 7 < 1.

End

3.5 Applications to 2D Corpus Callosum Shapes

We applied the algorithm to nine sets of 2D corpus callosum slices. The fea-

ture points were extracted with the help of a neuroanatomical expert. Figure 3-3

shows the nine corpus callosum 2D images, labeled CC1 through CC9. In our

experiments, we first did the simultaneous clustering and matching with the corpus

callosum point sets CC5 and CC9. The clustering of the two point sets is shown

in Figure 3-4. There are 68 cluster centers. The circles represent the centers and

the dots are the data points. The two cluster centers induce the diffeomorphic

mapping of the 2D space. The warping of the 2D grid under this diffeomorphism

is shown in Figure 3-5. Using this diffeomorphism, we calculated the after-image

of original data points and compared them with the target data points. Due to the









large number of cluster centers, the cluster centers nearly coincide with the original

data points and the warping of the original data points is not shown in the figure.

The correspondences (at the cluster level) are shown in Figure 3-6. The algorithm

allows us to simultaneously obtain the diffeomorphism and the correspondence.

Using our formulation, we are able to calculate the geodesic distances between

the two sets of cluster centers. This is done on the shape manifold. Each point q on

the shape manifold M represents a set of N cluster centers xa, x2, ... XN E I2 and

has a coordinate q (x{, X, x X,..., X x2) where xi = (x, ), i 1, 2,..., N.

Let q(t) be the geodesic path connecting two points ql and q2 on the manifold.

Using the norm defined for the tangent vector in Camion and Younes [ ], the

geodesic distance between qg and q2 is


Dgeodesic (q, q2) j TQ2(Q K(q)Q2 + A)-1QTdt. (3.21)

where K(q) is the kernel of the L operator evaluated at q(t) and as mentioned pre-

viously, the thin-plate spline kernel is used. Q2 comes from the QR decomposition

of P


R
P (Q1: Q) (3.22)
0

and



P = (3.23)



\ x xN
We also experimented with different number of cluster centers. Table 3-1

shows a modified Hausdorff distance as first introduced [ ] between the image









set of points CC5 after diffeomorphism and the target set of points CC9 when

the number of clusters vary. The reason for using the modified Hausdorff distance

instead of the Hausdorff distance is that the latter is too sensitive to outliers. The

definition of the modified Hausdorff distance is



Hmod(A, B) = max(hmod(A, B), hmod(B, A)), (3.24)

where A and B are finite point sets and



hmod(A,B) min a b (3.25)
aeA
is the average of the minimum distances instead of the maximum of the mini-

mum distances. It is easy to see that when the number of clusters increases, the

matching improves as the modified Hausdorff distance decreases.

In the third column in Table 3-2, we list the geodesic distances between the

two sets of cluster centers after pair-wise warping and clustering all the pair of

corpus callosum point sets. Using the cluster centers as landmarks, a diffeomorphic

mapping of the space is induced. With this induced diffeomorphism, we mapped

the original data sets and compared the image of the original point set under the

diffeomorphism and the target point set using the modified Hausdorff distance. The

modified Hausdorff distances between the pairs are listed in the fourth column in

Table 3-2. Finally, from the original nine corpus callosum point sets, we warped

the first eight point sets onto the ninth set and Figure 3-7 displays the overlay of

all point sets after diffeomorphic warping.

3.6 Applications to 3D Shapes

We applied our formulation and algorithm to the 3D point data of hippocam-

pal shapes. We first applied the algorithm to synthetic data, where we have the













:'."* 0.2 02

..0 0

-0.2 -0.2
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2





0.2 .* 0.2

0 0 ..

-0.2 -0.2
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2





0.2 0.2
..-; ../** : o.2 : ":;' ..8 .o.2 .
'$ 0 *** 0 t::

-0.2 -0.2
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1 1.2

Figure 3-3: Point sets of nine corpus callosum images.


0.5
0.5
0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1 9.

00-
-0.1
-0.1
-0.2
-0.2
-0.3
0.4 0.6 0.8 1 1.2 0.4 0.6 0.8 1

Figure 3-4: Clustering of the two point sets.


0.4- 0.4
0.3- 0.3
0.2- 0.2
0.1 0.1
0-- 0
-0.1--- -0.1
-0.2--- -0.2

0.2 0.4 0.6 0.8 1 1.2 0.2 0.4 0.6 0.8

Figure 3-5: Diffeomorphic mapping of the space.


1 1.2








57


0.4



0.3 OQ 0
0.25 -0 0 0 0 o 0 @.OO
0.15 o; o a coo o^:




0.3 00 0 o 0- O o. .:o ... 00,- ..
S 0 : '



S..o



0.05- : o O 0

0












0.25 -





o
0.05 -
-0.05

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2












Figure 3-7: Overlay of the after-images of eight point sets with the ninth set.



Table 3-1: Modified Hausdorff distance of the matching point sets.

Number of Clusters Modified Hausdorff Distance
0.10 30.02













20 0.0082
30 0.0057
0.15 -.. ::b C) (O '' 06
























40 0.0050
50 0.0043
0.1 6: O 0















0.0568 0.0027
-0.1
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2





Table 31F Modified Hausdorff distance of the matching point sets.












68 0.0027











Table 3-2: Geodesic distances between two sets of cluster centers and modified
Hausdorff distances of matching points.


From To Geodesic M.H.
CC1 CC2 0.0264 0.0055
CC1 CC3 0.0132 0.0014
CC1 CC4 0.0289 0.0048
CC1 CC5 0.0269 0.0056
CC1 CC6 0.0250 0.0097
CC1 CC7 0.0323 0.0054
CC1 CC8 0.0256 0.0043
CC1 CC9 0.0241 0.0041
CC2 CC3 0.0277 0.0059
CC2 CC4 0.0342 0.0063
CC2 CC5 0.0308 0.0057
CC2 CC6 0.0211 0.0100
CC2 CC7 0.0215 0.0040
CC2 CC8 0.0271 0.0044
CC2 CC9 0.0258 0.0093
CC3 CC4 0.0443 0.0059
CC3 CC5 0.0294 0.0047
CC3 CC6 0.0181 0.0032
CC3 CC7 0.0256 0.0060
CC3 CC8 0.0153 0.0018
CC3 CC9 0.0305 0.0044
CC4 CC5 0.0231 0.0046
CC4 CC6 0.0304 0.0056
CC4 CC7 0.0324 0.0056
CC4 CC8 0.0311 0.0054
CC4 CC9 0.0434 0.0090
CC5 CC6 0.0266 0.0056
CC5 CC7 0.0325 0.0053
CC5 CC8 0.0225 0.0037
CC5 CC9 0.0305 0.0069
CC6 CC7 0.0244 0.0050
CC6 CC8 0.0186 0.0026
CC6 CC9 0.0274 0.0056
CC7 CC8 0.0212 0.0044
CC7 CC9 0.0241 0.0102
CC8 CC9 0.0196 0.0050







59




240 +
+, +,++

220 + + +
n + + + +t

+ + 4+ +._ + + -
200 +-+ +
S+ +++ + t_ + ...
8 ++ + + + + ++

+180+S++- +++
4.-. + *+ -W .+*
+I + c+ u -- + + t #

1++ +-+ + ..+ + + + ,..
+4+r+ + 4W + ++ + 4+ +++ + ...
180 + + +#. ++ + + .. :




*. -.. .
140


120
200 220 240 260 280 300 320 340 360

Figure 3-8: Two point sets of hippocampal shapes. The set with crosses is the
original set and the set with dots is the one after GRBF warping.


knowledge of ground truth and this serves as the validation of the algorithm. We

then experimented with real data and evaluated the results using various measures.

3.6.1 Experiments on Synthetic Data

We selected one hippocampal point set and warped it with a known diffeomor-

phism using the Gaussian Radial Basis Function (GRBF) kernel. We choose a = 60

for the GRBF because with this large value of a ,we are able to generate a more

global warping.

Figure 3-8 shows the two point sets of hippocampal shapes. The set with

crosses is the original set and the set with dots is the one after GRBF warping.

First, we have no noise added. We used the TPS kernel to recover the diffeo-

morphism via joint clustering using our algorithm. The reason we use different

kernels for warping and recovering is the objectiveness. It is trivial to recover the










Table 3-3: Matching errors on synthetic data with different noise levels

Noise level\No. clusters 100 200 300 400 500
0 0.21 0.17 0.16 0.19 0.20
0.1 0.300.13 0.280.09 0.260.08 0.290.09 0.310.11
0.2 0.410.16 0.390.12 0.350.11 0.370.13 0.390.14
0.3 0.440.17 0.410.13 0.390.15 0.400.16 0.420.19
0.4 0.610.23 0.540.19 0.520.18 0.550.20 0.590.21
0.5 0.680.24 0.620.25 0.590.24 0.630.22 0.650.28
0.6 0.820.38 0.750.35 0.720.33 0.760.37 0.800.36
0.7 0.960.49 0.900.42 0.860.42 0.900.44 0.940.46
0.8 1.210.54 1.130.51 0.920.48 1.090.49 1.180.51
0.9 1.630.72 1.480.66 1.450.62 1.490.61 1.520.68
1.0 1.820.78 1.700.71 1.640.67 1.690.73 1.770.75


deformation that is warped with the same kernel. Since the reference data are

synthesized, we know the ground truth and we are able to compare our result with

the ground truth. After unwarping the point set with our recovered diffeomor-

phism, we find the squared distances between the corresponding data points, and

find the average and then take the square root. This is the standard error for our

recovered diffeomorphism. We have two free parameters, A and Tfinal. Tfinal is

determined by the limiting value of TT which is in turn determined by the number

of clusters. We choose a A value such that the whole optimization process is stable

in the temperature range from initial T to Tfinal. We experimented with different

numbers of clusters and listed the corresponding standard errors in the first row

of Table 3-3. It is easy to see that the standard error goes down as the number of

clusters goes up from 100 to 300 and goes up again when the number of clusters

increases further. This is because when we have too few clusters, the points are not

well represented by the cluster centers. On the other hand, if we have too many

clusters, the variance between the two shapes is too big and the deformation in-

creases dramatically. For the standard error, there is an optimal number of clusters

and in this case we find it to be 300. We need to estimate the macroscopic and

microscopic dimensions of the shape in order to see how big the standard error

is. We calculated the covariance matrix of the original data set. We find their









Table 3-4: Limiting value of a determined by the number of clusters

Number of clusters 100 200 300 400 500
Limit a 3.3 1.9 1.2 0.9 0.6


eigenvalues to be [48.1, 11.8, 4.1]. This gives us an estimate of the macroscopic

dimensions to be about 100, 24 and 8, namely twice the eigenvalues. We then find

out the average distance between the nearest neighbors to be 2.65. This give us the

microscopic dimension of the shape. As we can see from the table, our matching is

very accurate.

Next we add noise to the warped data and test the robustness of our algorithm

to noise. After GRBF warping, we add Gaussian noise to the warped data with

different variances a. We experimented with ten trials for each noise level from

0.1 to 1.0 and for each cluster level from 100 to 500. The standard errors and

the deviation are shown in Table 3-3. We can see the standard error increase

with the increasing noise level but it approximately -Il. -, in the range of the

noise. Stronger noise does not increase the matching error dramatically and this

shows the algorithm is robust against noise. This is easier to see when plotted

in Figure 3-9 with error bars. We split the five levels of clusters into two plots

because it looks messy if they were put together in a single plot Figure 3-9(a) has

the errors for 100, 200 and 300 clusters and Figure 3-9(b) has the errors for 300,

400 and 500 clusters. We can see that at the 300 cluster level, we obtain the best

matching.

3.6.2 Experiments on Real Data

We applied the algorithm on different real hippocampal data sets. Figure 3-10

shows two hippocampal shapes. Figure 3-11 shows the annealing process. The

x axis is the iteration step. The dashed line is the scaled temperature VT/2 or

l(T. The solid line is the actual variance a. We can see when the temperature

goes down, it drives the a down. We observe a phase transition at temperature









T = 3.37 x 103. We observe that there is a lower limit for a. When the temperature

gets very low, the a becomes a constant, which is 1.2, and no matter how much

lower the temperature gets, the a -I. -i, constant. This constant is determined

by the number of clusters. In Figure 3-12a through Figure 3-12e, we show how

this limit changes with the number of clusters. When the number of clusters

equals or exceeds the number of data points, the limit approaches zero. Table 3-4

displays the limits of a changing with the number of clusters. Because of noise and

sampling error, we should not allow this limit to go to zero. Again we observe when

we have 300 clusters, we have a reasonable a = 1.2 as we recall the average distance

between the nearest neighbors is about 2.65.

Figure 3-13 shows the clustering of the two shapes. We then did the matching

for all the pairs out of ten hippocampal shapes. Table 3-5 and Table 3-6 shows

three measures for the matching results with different clusters: Jensen-Shannon

divergence, Hausdorff distance and modified Hausdorff distance.

The Jensen-Shannon divergence (a special case with A = 1/2) is defined as [

D (p(x)log + q(x) log x )dx, (3.26)
D n ( p(x) + q(x) p(x) + (x)

where x is the random variable while p(x) and q(x) are the two probability

densities. Notice this measure is highly non-linear. When p(x) and q(x) are

completely independent, namely in our matching case, when the two shapes are

completely different, D has a maximum of 2 log 2 = 1.39. In practice, we use

the following technique to compute D. We observe in (3.26), the integral can

be expressed as the expectation values of some functions under two different

probability distributions:


2p(x) 2q(x)
D < log p(x > + < log > (3.27)
p(x) + q(x) p(x) + (x)









where logp(x(x) > is the expectation value of function log p( x) under

probability distribution p(x) and < log 2q(x) ) >q is the expectation value of

function log 2q(x) ) under probability distribution q(x). In our Gaussian mixture

model, we see the data points as samples from a Gaussian mixture probability

distribution with known cluster centers. Here


p(x) = N (223/2 X 2 k 2 (3.28)
N (2-2)3/ exp( x2T 2
k-1
and
1 1 N ||x |2
q(x) N2 exp( IIX Si (3.29)
N (27rT2)3/2 ex 2(2
k-1
where x is the random variable, namely the space location and {rk} is the first set

of cluster centers and {sk} is the second set of cluster centers. We use the average

of finite samples as an approximation of the expectation values and we have


1 2p(x) 1 2q(yj) (3.30)
D > log P log (3.30)
N I Aj p(xi) q(xi) N2 1 p(yj) ()

where {xi} is the first set of data points; N1 is the number of points in the first set;

{yj} is the second set of data points; and N2 is the number of points in the second
set. We have seen that Jensen-Shannon divergence is very useful in estimating the

validity of the registration of two point shapes without knowing the ground truth.

The Hausdorff distance is defined as


H(A, B) = max(h(A, B), h(B, A)), (3.31)

where A and B are finite point sets and


h(A, B) maxmin I a b | (3.32)
acA bcB

The Hausdorff distance measures the worst case difference between the two

point sets. From Table 3-5 and Table 3-6 we can see that when we have 300







64

clusters, we have the minimum Jensen-Shannon divergence and the Hausdorff

distance. However, the Hausdorff distance is too sensitive to outliers. We also

calculated the modified Hausdorff distance as first introduced in Dubuisson and

Jain [ ]. The definition of the modified Hausdorff distance was given before

in (3.24) and (3.25). It is the average of the minimum distances instead of the

maximum of the minimum distances. It is easy to see that when the number of

clusters increases, the modified Hausdorff distance decreases.























A N=100
N N=200
2.5- 0 N=300



2



1.5


0 02 04 06 08


0 N=500
N=400
2.5 0 N=300


, jtI jiI


0 0.2 0.4 0.6 0.8


Figure 3-9: Matching errors on synthetic data for different number of clusters


o, 0
















































'....


. *


280


260


Figure 3-10: Two hippocampal shapes









67



















70
x100


60



50-



40- 0



30



20



10



0
0 10 20 30 40 50 60 70 80 90 100
iteration steps


Figure 3-11: Deterministic annealing in the clustering process: the dashed line is
the scaled temperature VT-/2 or TT-. The solid line is the actual variance u. When
the temperature goes down, it drives the a down. There is a phase transition at
temperature T 3.37 x 103 and there exists a lower limit 1.2 for a.





















iteration steps

(a) 100 clusters










iteration steps

(b) 200 clusters










Iteratlon steps

(c) 300 clusters










Iteratlon steps

(d) 400 clusters
\































iteration steps

(e) 500 clusters

Figure 3-12: Limiting value of determined by the number of clusters
Figure 3-12: Limiting value of ar determined by the number of clusters








69















15 .


0 i 04 0





160 22







(a) Clustering of the template hippocampal shape










10 ..
1340














2 40


1 260
-- 2220






















(b) Clustering of the referencmplate hippocampal shape
210.1 .?(




20140 2. 0 0



12D ^ 240








(b) Clustering of the reference hippocampal shape


Figure 3-13: Clustering of the two hippocampal shapes












Table 3-5: Jensen-Shannon divergence for various pairs of shapes


Jensen-Shannon div.
Trial no.\No. 100 200 300 400 500
clusters
1 0.87 0.31 0.03 0.13 0.21
2 0.93 0.62 0.47 0.05 0.24
3 0.76 0.27 0.04 0.16 0.32
4 0.98 0.52 0.34 0.09 0.45
5 0.69 0.41 0.14 0.18 0.36
6 0.57 0.23 0.43 0.78 0.97
7 0.66 0.21 0.05 0.14 0.30
8 0.99 0.70 0.25 0.19 0.63
9 0.85 0.42 0.11 0.68 0.74
10 0.97 0.62 0.10 0.18 0.55
11 0.70 0.33 0.06 0.13 0.26
12 1.02 0.64 0.08 0.44 0.71
13 0.89 0.54 0.20 0.31 0.65
14 0.57 0.09 0.15 0.66 0.80
15 0.88 0.30 0.05 0.29 0.36
16 0.90 0.75 0.12 0.17 0.44
17 0.61 0.16 0.28 0.53 0.72
18 0.91 0.37 0.18 0.40 0.88
19 1.12 0.80 0.47 0.09 0.28
20 0.96 0.54 0.33 0.60 0.74
21 0.65 0.23 0.51 0.78 1.04
22 0.93 0.46 0.22 0.51 0.68
23 0.92 0.60 0.28 0.15 0.34
24 0.80 0.26 0.57 0.69 0.86
25 1.10 0.62 0.44 0.78 0.97
26 0.90 0.39 0.05 0.21 0.47
27 0.58 0.07 0.20 0.56 0.77
28 0.93 0.51 0.09 0.40 0.63
29 0.99 0.26 0.18 0.37 0.70
30 0.60 0.06 0.17 0.54 0.57
31 0.83 0.19 0.08 0.37 0.76
32 1.22 0.42 0.57 0.70 0.95
33 0.80 0.59 0.30 0.86 0.92
34 0.89 0.76 0.35 0.28 0.67
35 1.05 0.42 0.37 0.81 1.13
36 0.92 0.25 0.31 0.60 0.85
37 0.79 0.35 0.08 0.24 0.40
38 0.90 0.42 0.16 0.35 0.68
39 0.86 0.27 0.38 0.50 0.71
40 0.55 0.04 0.19 0.36 0.67
41 1.02 0.30 0.47 0.81 0.98
42 0.43 0.07 0.22 0.56 0.86
43 0.78 0.56 0.18 0.39 0.61
44 0.61 0.09 0.25 0.70 0.82
45 0.44 0.15 0.26 0.53 0.93












Table 3-6: Hausdorff and modified Hausdorff distance for various pairs of shapes

Hausdorff distance modified Hausdorff
Trial no.\No. 100 200 300 400 500 100 200 300 400 500
clusters
1 7.1 7.4 5.7 6.2 7.3 2.8 2.0 1.4 1.2 1.1
2 9.3 8.9 7.2 8.3 8.7 3.5 3.1 2.8 2.4 2.3
3 7.2 6.1 4.9 5.6 6.4 2.0 1.7 1.4 1.3 1.2
4 8.4 7.8 7.2 5.2 6.5 2.7 2.4 2.3 1.7 1.4
5 9.6 9.7 8.0 8.4 8.9 3.9 3.6 3.1 2.8 2.7
6 9.2 6.3 7.1 7.8 8.6 3.1 2.8 2.5 2.2 2.1
7 6.9 5.8 4.4 6.0 7.3 2.4 2.2 2.1 1.7 1.5
8 8.9 8.5 7.0 6.4 8.2 3.0 2.6 2.4 2.2 1.9
9 9.3 8.0 5.9 7.6 9.1 2.9 2.7 2.3 2.1 1.6
10 7.8 7.3 4.7 6.7 8.1 3.2 2.8 2.3 1.8 1.4
11 8.7 7.7 5.8 7.4 9.0 2.5 2.1 1.6 1.4 1.2
12 9.1 8.3 6.1 7.4 8.6 3.3 3.0 2.5 2.2 2.0
13 9.3 8.4 6.5 7.0 8.7 3.6 3.4 3.1 2.4 2.2
14 7.4 5.1 5.5 6.8 8.3 3.0 2.7 2.3 2.0 1.8
15 8.8 7.1 4.9 6.3 7.8 2.6 2.0 1.5 1.3 1.2
16 9.4 9.0 6.1 7.3 8.5 3.2 3.0 2.4 2.1 1.9
17 8.6 6.8 7.9 9.0 9.9 3.4 3.5 3.1 2.7 2.4
18 9.5 8.2 6.5 7.4 8.0 3.7 3.2 2.9 2.6 2.1
19 9.2 7.8 7.2 5.1 6.4 2.8 2.6 2.3 2.2 2.0
20 9.6 8.0 7.3 8.7 9.3 3.9 3.5 3.3 3.1 2.7
21 8.4 6.1 6.9 7.8 9.5 3.3 3.1 2.8 2.4 2.3
22 9.7 8.5 7.0 8.1 9.0 2.9 2.7 2.6 2.3 2.1
23 9.6 8.2 7.3 6.5 7.7 2.4 2.1 1.7 1.6 1.5
24 7.8 6.6 7.2 8.9 9.6 3.1 2.7 2.5 2.2 1.9
25 9.8 7.9 7.6 8.8 9.2 3.8 3.4 3.0 2.8 2.7
26 9.0 7.3 5.8 7.0 8.7 2.9 2.4 2.0 1.4 1.2
27 7.8 6.0 6.5 7.2 8.3 3.2 2.9 2.5 2.1 1.9
28 9.5 8.1 6.1 7.4 8.8 3.0 2.8 2.5 2.1 1.8
29 9.7 8.3 6.7 7.0 8.5 3.4 2.7 2.2 1.9 1.7
30 7.1 5.6 6.2 7.3 7.8 2.5 2.2 1.8 1.3 1.2
31 9.3 6.7 5.8 7.5 8.4 2.7 2.4 2.2 1.6 1.4
32 9.9 7.2 7.8 8.5 9.2 3.5 3.0 2.8 2.5 2.3
33 8.9 8.0 6.3 7.1 8.7 3.3 3.1 2.6 2.4 2.3
34 8.9 8.5 6.7 6.4 7.0 3.0 2.8 2.5 2.2 2.1
35 9.6 8.4 7.0 8.1 9.9 3.7 3.4 3.0 2.3 2.0
36 8.7 5.9 6.2 7.5 8.4 3.2 2.9 2.5 2.3 2.2
37 7.7 6.3 5.4 6.1 7.6 2.4 2.2 1.9 1.6 1.5
38 9.5 7.1 5.7 6.6 7.9 2.8 2.4 2.3 2.0 1.8
39 9.2 7.3 8.2 8.4 9.0 3.2 3.0 2.7 2.5 2.4
40 6.5 5.2 5.8 6.7 7.3 2.0 1.6 1.3 1.1 0.9
41 9.4 7.5 7.9 8.6 9.2 3.8 3.2 2.5 2.4 2.3
42 7.2 5.8 6.7 7.3 7.9 2.3 1.9 1.6 1.5 1.4
43 8.2 7.4 6.0 6.7 7.8 2.9 2.4 2.2 2.0 1.8
44 7.1 5.9 6.6 7.5 8.0 2.2 1.9 1.7 1.5 1.4
45 7.0 6.5 7.2 7.7 9.1 2.5 2.1 1.9 1.7 1.6














CHAPTER 4
TOPOLOGICAL CLUSTERING AND MATCHING

In this chapter we extend our diffeomorphic point matching theory and

algorithm to include the known information of the topology of the underlying

shapes. In Section 4.1 we provide a brief introduction to the basic concepts

of topological spaces. In Section 4.2 we review the Kohonen Self-Organizing

Feature Map (SOFM) which was first introduced by Kohonen in the context of

neural networks in 1980s. Our topological clustering and matching is related

to SOFM because the essence of SOFM is topology preserving. However, our

topological clustering and matching is different from SOFM in many ways and

as part of Section 4.3 we discuss these differences. In Section 4.3 we discuss the

motivation and methods of topological clustering and matching. We do this with

graph topology assigned to the set of cluster centers. The graph topology can be

prescribed if we have prior knowledge of the shape topology or it can be arbitrary

in the case we don't know the shape topology in advance. In Section 4.5 we present

results for clustering and matching with prescribed topology with example of chain

topology, ring topology and genus zero closed surface topology, or S2 topology.

Section 4.6 describes how we can approximate the topology if we do not have the

prior knowledge about the topology of the shape in advance.

4.1 Fundamentals of Topological Spaces

The heart of topology is the concept of ii' ii ii. --," described by -i'. illor-

hood". We first introduce the concept of topological space. Readers can confer

the Encyclopedic Dictionary of Mathematics compiled by Mathematical Society of

Japan and translated by Massachusetts Institute of Technology [ ]. We also cite

from a book by Bourbaki [ ]and a book by Rosenfeld [ ].









Felix Hausdorff in his Foundations of Set TI.,,, y (Grundziige der Mengenlehre.

Leipzig, 1914) [ ]defined his concept of a topological space based on the four

axioms.

Let X be a set. A neighborhood system for X is a function it that assigns

to each point x of X, a family il(x) of subsets of X subject to the following axioms

(U):

(1I) x E U for each U in i(x).

(W2) If U1, U2 E i(x), then U1 n U2 E i(x).

(W3) If U E i(x) and U C V, then V E i(x).

(W4) For each U in it(x), there is a member W of it(x) such that U E it(y) for

each y in W.

it(x) is interpreted as the family of all neighborhoods of point x. An element

U E it(x) is called a neighborhood of point x. The intuitive translation of the above

axioms is as follows.

(W1) x is in each neighborhood of x.

(W2) The intersection of two neighborhoods of x is a neighborhood of x.

(W3) If a set V contains a neighborhood of x, then V is itself a neighborhood

of x.

(W4) For each neighborhood U of x, there is another neighborhood W of x,

such that U is the neighborhood of each point y in W.

Pavel Sergeevic Aleksandrov [Alexandroff] proposed in the paper On the

foundation of n-dimensional topology (Zur Begriindugn der n-dimensionalen

Topologie. Leipzig, 1925) [ ]:

A system of open sets for a set X is a family D of subsets of X -.I i- i,-;

the following axioms (O):

(01) X, 0 D.

(D2) If 01, 02 E D, then 01 n 02 E 0.









(03) If Ox E (A E A), then UAeA OA E 0.
The elements in D are called open sets. An easy intuitive interpretation of this

set of axioms is

(01) The empty set is an open set. The entire space X is an open set.

(02) The intersection of two open sets is an open set.

(03) The union of arbitrarily many open sets is an open set.

Using the set complement and DeMorgan's law we get a system of closed sets.

A system of closed sets for a space X is a family 3 of subsets of X satisfy-

ing the following axioms (F):

(31) X, 0 E 3.

(32) If F1, F2 e 3, then F, U F2 C .

(33) If FA E 3 (A c A), then AeA FA C &.
The elements in 3 are called closed sets. An easy intuitive interpretation of

this set of axioms is

(31) The entire space X is a closed set. The empty set is a closed set.

(32) The union of two closed sets is a closed set.

(53) The intersection of arbitrarily many closed sets is a closed set.
Kazimierz Kuratowski in the paper The operation A of analysis situs

(L'operation A de l'analysis situs. Warsaw, 1922) [ ] proposed:

A closure operator for a space X is a function that assigns to each subset A

of X, a subset Aa of X -. I i-fvi i-; the following axioms(C):

(TI) a 0.

(T2) (A U B) A U Ba.

(C3) A c AN.

(T4) A = A"

The intuition of the closure of a set A is the union of A and its boundary. This

set of axioms explained in words is as follows.









(1) The closure of the empty set is the empty set.

(2) The closure of (A U B) is the union of closure of A and the closure of B.

(3) The closure of A contains A as a subset.

(4) The closure of the closure of A is the same as the closure of A. This is

saying that the closure operator is idempotent.

Related to the closure operator is an interior operator.

An interior operator for a space X is a function that assigns to each subset

A of X a subset A' of X -.i i-fvi-; the following axioms (I):

(J1) Xi = X.

(J2) (A n B)' A n B'.

(J3) A' c A.

(J4) A" A'.

The intuition of the interior of a set A is A minus its boundary. This set of

axioms explained in words is as follows.

(J1) The interior of the entire space X is itself.

(32) The interior of (A n B) is the intersection of the interior of A and the

interior of B.

(33) The interior of A is a subset of A.

(34) The interior of the interior of A is the same as the interior of A. This is

saying that the interior operator is idempotent.

All these systems equivalently define the topological space. We can interpret

one easily in the language of another.

We can define open set using neighborhood:

A set A c X is an open set, if Vx E A, there exists a neighborhood U of x such

that U C A, then A is an open set.

Define closed set using open set:

A set A c X is a closed set, if X A is an open set.









Define closure of A using closed set:

If A C X, the closure of A is the set A" n= Bx, where Bx is a closed set and

A c B. In other words, the closure of A is the smallest closed set that contains A.

Define interior using closure:

Ai X (X A). In other word, for set A C X, first find the complement

of A, B =X A. Then find B", the closure of B. The the interior of A is the

complement of B".

Define neighborhood using interior:

A set U c X is the neighborhood of point x E X, if x E U', ,,in,,, li x is in the

interior of U.

Figure 4-1 shows some examples of different topological spaces.

Another importance concept in topological spaces is the separation axioms,

dictating the extent the points are separated from each other. Of our interest is one

particular space, called Hausdorff space.

Definition 4. A /i i',,,1,:1, ,,J space is called Hausdorff space if any two distinct

points have disjoint .:,in l .....i.i .

4.2 Kohonen Self-Organizing Feature Map (SOFM)

Kohonen developed the Self-Organizing Feature Map algorithm. He described

the SOFM in the context of neural networks with an aim to understand the the

brain cortex mapping of sensory organs, like retina [ ]. In the Kohonen

network, there are two layers of neurons, with the first layer the input and the

second layer the output. The output neurons are arranged in a rectangular grid.

Each input neuron is connected with each output neuron. Each output neuron

is associated with d-weights, with d being the number of the input neurons.

Kohonen describes a procedure of the initialization and update of the weights.

The important part of the algorithm is that each neuron on the output grid

has a neighborhood and when each neuron updates itself, the neurons in the





























Figure 4-1: Different topological spaces


neighborhood update themselves accordingly. This is described in many books as

"topologically ordered map", "topology preserving map" or "topographical map".

There are many misconceptions and misnomer here. First, the neural network is

not such a map but it simulates such a map because the codomain is a discrete

space, namely the rectangular grid. Second, "topographical map" is a misnomer for

topologicall map." Third, what the network simulates is a continuous map. That is

the exact interpretation of the idea that "if the two outputs are in a neighborhood

then the two inputs are also in a neighborhood." "Topology preserving" is not

guaranteed. In fact this is obvious when the input domain is two dimensional and

the output domain is one dimensional. What we get is a plain filling Peano curve.

As it is well known that a plain region cannot be topologically equivalent to a curve

segment and the Peano curve is a continuous map but not a topological map, or

homeomorphism.

Kohonen only gives a procedure but does not give an objective function that

this procedure optimize. People soon find an interpretation of SOFM in the context









of clustering and dimension reduction of data, with mathematical abstraction.

Ritter et al. give an objective function [ ]



E{w} = i,, A(i, k)|lk i 2,
ik/p
where 1.,, is the membership matrix element, equal to 1 if patter is in cluster k

and 0 otherwise. A(i, k) is the neighborhood function between cluster i and cluster

k. A(i, k) = 1 if i = k and falls off with distance ||w w Wkl|. A typical choice for

A(i, k) is



A(i,k)=e

The Kohonen procedure is actually the clustering process of high dimensional

data. It tries to model the dimensionality reduction. The data points are in h-

dimensional Euclidean space but they may approximately lie on a two dimensional

manifold. The clusters are constrained with a 2D rectangular grid. During the

clustering it is required that neighboring cluster centers stay close to each other.

Thus this procedure provides an approximate discrete 2D patch which is a map

from a 2D rectangular region with grid to the 2D manifold embedded in h-

dimensional Euclidean space, where the data dwell. This provides an intrinsic

coordination (2D) for the data and it extracts 2D features from the data. Because

of this clustering process, the coordinates are only approximate and the coordinates

are discrete (the i, j indices). In the general case, each data point has h coordinate

components and thus live in Rh but the set of data only populate a 1-dimensional

submanifold M' of Rh. Suppose we have a k-dimensional grid of cluster centers,

U C Rk. Only when k = 1 can we have a homeomorphism from U to an open

neighborhood of M1. One example is that 1 = 3 and k = 2. If the embedded

manifold is roughly a 2D sheet with some non-negligible thickness in the third









dimension, the clustering result in an approximation in which we represent this 3D

curved thick sheet with a 2D thick-less sheet. If instead of a thick sheet, we have

a solid 3D manifold and we still want to use a 2D sheet to approximate it, it will

result in that the 2D sheet will wrinkle and scramble so that it tries to fill the 3D

space. It is easier to visualize when I =2 and k = 1 and then we have a space

filling Peano curve, resulting a continuous map but not an homeomorphism.

4.3 Topological Clustering and Matching

4.3.1 Why: the Need for Topology

When the two shape differ by a large deformation, complications do occur.

Figure 4-2 shows the contours of two hand shapes. The hand on the left has

the thumb and the fore finger pointing out while the hand on the right has the

fore finger folded. These two shapes differ by a large deformation. We apply our

diffeomorphic point matching algorithm to these two shapes and Figure 4-3 shows

the clustering of the two shapes. While the clustering looks pretty good, a closer

examination of the correspondence as shown in Figure 4-4 indicates that there are

incorrect correspondences. We have discussed in Chapter 1 the issues that there

is no way to clearly define what the correct correspondence is, with our visual

intuition, we know this is not the correspondence we want. This is the case where

nearby points correspond to points that are not nearby, a violation of topological

property of the mapping. We want to enforce the constraint that nearby points are

mapped to nearby points and introduce the topology constraint to the matching.

This has significance in two scenarios:

1. The problem definition of the shape matching does not have topology

constraint and the objective function does not have the topology term.

However, the numerical procedure of solving this problem may be caught

up in a local minimum, which gives incorrect correspondence. In this case,






















Figure 4-2: Image contours of two hands


introducing the topological constraint will help avoid the local minimum and

find the correct correspondence as defined by the objective function.

2. The numerical procedure does find the global minimum defined by the ob-

jective function without topological constraint. However, the correspondence

is still not what we intended. In this case, the point set as a set of point

without any topological structure is not sufficient to describe the underlying

shape. For example of the two hand shapes as in Figure 4-2, we know that

the points lie on a contour curve. By adding this topology requirement, we

are actually defining a different problem from that without topology structure

and of course the solutions of the correspondence should be different with and

without topology since they are two different problems.

4.3.2 How: Graph Topology

It is clear that the information of the topological structure of the underlying

shape of the point set helps define and solve for the correct correspondence. Then

it is natural to introduce some topological structure in the set of cluster centers and

a graph is the easiest way to represent this topology. We make the cluster centers

the vertices of the graph and assign edges between the vertices.

An intuitive thinking of the topology construction on the graph may be

that make the set of vertices the support set and make the .,li..:'ent vertices
















-50 H


0 0
O O



O f
JO

S


100 150 200 250


Figure 4-3: Clustering of two hands


Figure 4-4: Correspondence


50 H


0
o



0

b3


9nn k


O
O
00
O O
O i,
o
o
o

og
o
o
o
o












U2



B UA

TU A
U3




Figure 4-5: Finite topology

neighborhood. A more careful analysis will show that this idea should fail. Since

the set of the vertices is a finite set, we prove a property of finite topology.

Theorem 7. Let (X, T) be a l,.1../.. .,i space and X is a finite set. T is Hausdorff

iff T is discrete topology.

Proof. First we prove if T is discrete topology then T is Hausdorff.

The proof is trivial because in discrete topology every singleton set of a point

is an open set. For any two distinct points A E X and B E X, the di- i.iil

neighborhood UA and UB are UA = {A} and UB {B}.

Next we prove if T is Hausdorff then T is discrete topology.

We use proof by contradiction. Now we assume the contrary, namely T is

Hausdorff but T is a topology other than discrete topology. Then there must exist

a point A E X, such that {A} is not an open set. Let UA be the smallest open set

that contains A:



UA= Un U n... n k,

where U1, U2,..., Uk are all the open sets that contains A. By assumption, UA must

have at least another point different from A. We call this point B as shown in









Figure 4-5. Now it is obvious that there is no open neighborhood of A that does

not contain B. Hence T is not Hausdorff. So the original claim is proved.



Since the Euclidean space is a Hausdorff space and the shapes we consider

as sub-topological spaces of the Euclidean space are Hausdorff spaces, we are not

interested in graph topologies that are not Hausdorff. However, from the above

theorem we know that the only Hausdorff topology is discrete topology. The

discrete topology is not good here because each point is an open set and each point

can have a open set consisting of itself. Each point is completely discrete, meaning

isolated and disconnected. So the discrete topology, in some sense, is no topology,

no 'ii .ii i. -C" or topological structure.

The correct approach is through in 1,gi'p realization or qintili< embedding. This

is the study of a branch of graph theory, ,.,, '.1..;'''''' i1 rtip theory [ Intuitively,

graph realization or graph embedding is to think the vertices of the graph as

points in an Euclidean space and the edges of the graphs as the lines or curves in

the Euclidean space connecting the vertices. With the graph realization or graph

emb.1.1iir.-. the graph G is a sub-topological space of the shape S as a topological

space. So if the points in the graph G are in a neighborhood, then they are also in

a neighborhood in the shape topological space S.

In the following of this chapter, we develop topological clustering techniques

for the purpose of diffeomorphic point matching. It has many similarities with

Kohonen SOFM but there are also many differences. First, the purpose of SOFM

is dimensionality reduction while our topological clustering is for point matching.

Second, the map in SOFM is local while our topology is global. SOFM only

provides a single patch for one open neighborhood of a point on the submanifold

while our clustering allows non-trivial topologies which cannot be covered by a

single patch. Third, the interaction in SOFM is between one cluster center and









the data points in the neighboring clusters while the interaction in our model in

between one cluster center and the neighboring cluster centers.

4.4 Objective Function and the Algorithm

In order to enforce the principle of that points in a neighborhood should stay

in the neighborhood, we add the term

N N
G. Jmn Frm r112 (4.1)
m= in=1

to the objective function, where G is the .,li. 1:ency matrix of the graph of the

cluster centers of one set. The other set has the similar topological constraints.

Gmn is 1 if there is an edge between rm and r,, and 0 otherwise. The new objective

function now is


E(\ ',. ,r,s, v, ) (4.2)
N1 N Ni N
Su Xi rk\ 2 + 24,72 Y \I log\I
i=l k=l i=l k=1
N N2 N
+ E E i-S +2E Ek
k=l j=1 k=1
N 1
Y, I 8k -rk 1)2 2T 2A] ] Lv(x,t) 2dxdt
k=1
N N N N
y >1-0 n F F .2 + Tp 12
S i q lq 1 1P 11
m=l n1=l q=l q=l

The matrix G is a symmetric matrix. The reason for the factor in the last line

is that each edge in the graph is counted twice in the summation. T is a new

parameter describing the strength of the links between the cluster centers.

The algorithm to minimize this energy is very similar to the algorithm

introduced in Chapter 3. However, the update equations for the cluster centers

(3.19) and (3.20) should be modified accordingly:











z1- ui; E71 fo aj(t)K(qj(t), Qk(t))dt +' Em- G1 k m
1 i-l1' +T Tn I Zp


EY1 Mjkyj + (rk, 1) + T Em 1m Grnk rn
Sk = ---- y------7-- Vki. (4.4)
1 IMjk + T i. m Grnk
The rule of modification is, when updating the cluster center positions for rk,

consider all other cluster centers r,, m = 1, 2,..., N in the graph, whenever there is

an edge from r, to rk, we add Trr, to the numerator and we add T to the

denominator. The modification to the update of Sk in the second set is similar.

4.5 Prescribed Topology

In some situations, the class of shapes have the same topology and the

topology is known and the graph representation of the topology is easy. In such

cases, we can initialize the graph with the prescribed topology. In the 2D case,

when we deal with line contours, two typical situations are the open curves and

closed curves. We can use chain topology and ring topology for the graphs.

4.5.1 Chain Topology

We solved the matching problem again with the hand shapes with chain

topology. The clustering result is shown in Figure 4-6 and this way we find the

correct correspondence in Figure 4-7.

4.5.2 Ring Topology

We applied the topological clustering and matching algorithm to the corpus

callosum data, with ring topology. Figure 4-8 shows the topological clustering

while Figure 4-9 shows the correspondence.

4.5.3 S2 Topology

We know the hippocampus shapes have a S2 topology. We initialize the graph

as a latitude and longitude grid. Figure 4-11 is the topological clustering of the





























000

0
c D~O


0 50 100 100 150 200 250


Figure 4-6: Topological clustering and matching of two hands


50 100 150 200


Figure 4-7: Correspondence with topology constraint


-50 k


50 H


9nn C
























-0.2


0.4 0.6 0.8 1 1.2


0.4 0.6 0.8 1


Figure 4-8: Topological clustering of corpus callosum shapes


-0.2 [


&OIRO


GMWIIV































































0.5 0.6 0.7


0.8 0.9


Figure 4-9: Correspondence in topological clustering of corpus callosum shapes

















9

8

7

6

5

4

196

194 296
294
192 292
292




Figure 4-10: Sphere topology


first hippocampus set and Figure 4-12 is the topological clustering of the second

hippocampus set.

4.6 Arbitrary Topology

We can see the limitations with prescribed topology. First, the class of

shapes may have different topologies. Second, even if the class of the shapes have

the same topology, the topology may be unknown before we run the matching

algorithm. Third, even if the topology is known, it may be too complicated

to construct a graph approximation and the initialization may involve human

intervention. Forth, the edges of the graph is not truly topological in a sense they

are not indefinitely flexible and stretchable strings. The topological constraints

in the objective function 4.2 are actually identical elastic strings. So we see the

geometrical or metric factor in the constraints. Sure we can adjust the coefficient