BMIRC主催セミナー 題目：広島大学工学部における医用画像処理の取り組み：放射線科，眼科，内科 講師：広島大学　准教授　玉木徹 日時：平成24年9月28日(金)13:00～14:00 主催：BMIRC http://www.bmirc.kyutech.ac.jp

1. h"p://kids.wanpug.com/top_person.html

2. h"p://www.sciencekids.co.nz/pictures/humanbody/braintomography.html

3. h"p://www.sciencekids.co.nz/pictures/humanbody/heartsurfaceanatomy.html

4. h"p://sozai.rash.jp/medical/p/000154.htmlA
h"p://sozai.rash.jp/medical/p/000152.htmlA
h"p://medical.toykikaku.com/ / /A
h"p://www.sciencekids.co.nz/pictures/humanbody/humanorgans.html

5. • CT
•
•
• NBI
h"p://sozai.rash.jp/medical/p/000154.htmlA
h"p://sozai.rash.jp/medical/p/000152.htmlA
h"p://medical.toykikaku.com/ / /

6. Non-rigid Registration for Medical Images
Using a Free-form Deformation
with Multiple grids
ToruAHigaki,AKazufumiAKaneda,AToruATamaki,ANobutakaADate,AShogoAAzemotoA:A"NonMrigidAImageARegistraPonAforAMedicalADiagnosisAUsingAFreeMformA
DeformaPonAwithAMulPpleAGrids,"A ,AVol.37,ANo.3,App.286M292A(2008A05).A
ToruAHigaki,AToruATamaki,AKazufumiAKaneda,ANobutadaADate,AShogoAAzemoto:A"NonMrigidAImageARegistraPonAforAMedicalAImagingAusingAaAFreeMformA
DeformaPon"AProc.AofAIEVC2007;AImageAElectronicsAandAVisualACompuPngAWorkshop,AInsPtuteAofAImageAElectronicsAEngineersAofAJapanA(2007A11).A
6

7. Medical Imaging Technology
• Growth of medical imaging devices
• Diagnosis using medical images
– Visualization
– Computer Aided Diagnosis
ImportanceAofAA
MedicalAImagingA
TechnologiesA
7

8. Diagnosis using multi-modal images
• Modality
– Medical imaging Devices
CT, MRI, PET, etc...
• Features of each modality
CT X-Ray Structure
MRI H atoms Tissue
PET FDG-tracer Cancers
ObservaPonAusingAmulPMmodalAimagesA
ProvidesAmoreAinformaPon!A
8

9. Alignment of Multi modal images
• Superimpose display
– CT + PET
Defining the locations of the cancers
+A =A
CTAimageA PETAimageA
9

10. Proposed deformation model
• Multiple control grid:
– Global grid
• entire image
• rough alignment
– Local grid
• observation area
• accurate alignment
14

11. Interaction between
global and local control grids
• Sequential operation
<Step1>
adjusting a global grid
align the global area
registraPonA
<Step2>
adjusting a local grid
align the local area registraPonA
15

12. Sampled images
ImagesA
CT (mono-modal)
Modality
taken at different times
Resolution 152×200
Observation
79×94
area
ProposedA
ControlAgridsA
Proposed B-Spline
Control Global 6×6
11×11
grid Local 6×6
order 5 3
BMSplineAbasedAFFDA
17

13. CT
Automatic Heart Segmentation in CT images
by Using Atlas
,A ,ABisserARaytchev,A ,A :A 62
, ,A 2011A10/22 .A
,A ,ABisserARaytchev,A ,A :A CT ,A
CAD 141 ,A ,A (2010A11).A
, ,ABisserARaytchev,A ,A :A ,AMIRU2010A
,App.894M897,A ,A (2010A07).A

14. ( )
2

15. 5 23
h"p://www.mhlw.go.jp/toukei/saikin/hw/jinkou/geppo/nengai11/kekka03.html#k3_2

16. CT
(endo) (WH), (epiLV)
, (RV) 4

17. CT SPECT
CT SPECT

18. SingleMphotonAemissionAcomputedA
tomographyA(SPECT)A XMrayAcomputedAtomographyA(CT)A
h"p://en.wikipedia.org/wiki/File:SPECT_CT.JPG h"p://en.wikipedia.org/wiki/File:Rosies_ct_scan.jpg
SIMENS,ASymbiaATASeriesASPECT•CTA
h"p://healthcare.siemens.com/molecularM
imaging/spectMandMspectMct/symbiaMt

19. GOAL
CT 4
CT WH epiLV endoLV RV
Idea
CT

20. • CT WH epiLV endoLV RV
•

21. • 6 1 5
6
Patient 1 Patient 2 Patient 3
Patient 4 Patient 5 Patient 6

22. VS.
1 2 3 4 5 6 1 2 3 4 5 6
Patient Patient
1 1
2 2
3 3
4 4
5 5
6 6

23. VS.
1 2 3 4 5 6 1 2 3 4 5 6
Patient Patient
1 1
2 2
3 3
4 4
5 5
6 6

24. • CT
•
•
• NBI
h"p://sozai.rash.jp/medical/p/000154.htmlA
h"p://sozai.rash.jp/medical/p/000152.htmlA
h"p://medical.toykikaku.com/ / /

25. Registration method for Cross-Sectional Images of
the Fundus considering Eye Movement
,A ,ABisserARaytchev,A ,A ,A ,A :A ,A ,Avol.A40,Ano.A4,A
pp.A609M614A(2011A07).A
,A ,ABisserARaytchev,A ,A ,A ,A :A ,A 246
,A09M02M01,pp.1M6,A ,A (2009A10).A
,A ,A ,A ,A ,A :A ,A D,A
Vol.J91MD,ANo.7,App.1808M1817A(2008A07).AA
,A ,A ,A ,A ,A ,A :A MIRU2007A
,App.487M492,A ,A (2007A07)A
,A ,A ,A ,A ,A ,A :A ,AVisualACompuPngA/A
CADA 2006A ,App.221M226,A ,A (2006A06).AA
,A ,A ,A :A OCT ,A ,AVol.34,ANo.4,App.
370M378A(2005A07).

26. h"p://www.sciencekids.co.nz/pictures/humanbody/eyediagram.html

27. ophthalmoscope
h"p://en.wikipedia.org/wiki/Ophthalmoscopy
h"p://medicalMdicPonary.thefreedicPonary.com/ophthalmoscope

28. Imaging modality
Optical Coherence Tomography (OCT)

29. Optical Coherence Tomography
Light source…
!
near-infrared light
Applying coherence of light
!
radial circular parallel Carl Zeiss Meditec Japan
OCT3000
scanning mode

30. Optical Coherence Tomography
1 OCT Time-domain OCT TD-OCT
1 [sec./ ]
10 [µ m]
2 OCT Spectral-domain OCT SD-OCT
0.01 [sec./ ]
5[µm]
Time-domain OCT

31. Goal
Time-domain OCT
Consideration:

32. [’07]
rotation
•
• 2 rotation
v
• u
Time-domain OCT
• OCT
•

33. RPE
18
512×256[pixel]
0[mm] 2[mm] (du,dv)
(d»1, d»2)
•
•

34. vs.
%

35. OCT
OCT 18
0 [deg.] 30 [deg.] 60 [deg.]
90 [deg.] 120 [deg.] 150 [deg.]
RPE
• SD-OCT
• Cup-to-Disc ratio C/D

36. Cup-to-Disc ratio (C/D )
ΦCup
CDR =
ΦCup
Φ Disc
ΦDisc
5 36
SD-OCT C/D

37. vs.
SD-OCT

38. OCT
1050 1600 2280

40. A method for generating retinal images for
investigation of the glare by intraocular lens
,ABisserARaytchev,A ,A ,A , ,A ,A
,AVol.A33,App.77M82(2012A06).A
,ABisserARaytchev,A ,A :A ,A 46 ,A14thATheAIRSJA2010, ,App.
48,A (2010A09).A
,A ,A ,A ,AB.ARaytchev,A ,A :A ,A ,A
(IOL&RS),AVol.A24,ANo.A1,A136M137A(2010A03).A
,A ,A ,ABisserARaytchev,A ,A ,A :A
,A 48 ,A 24 ,A 45 ,A22ndAAsiaAPacificAAssociaPonAofACataractA
andARefracPveASurgeonsAAnnualAMeePngA ,Ap.77,A ,A (2009A06).A
:A
MIRU2007A ,App.1171M1176,A ,A (2007A07)A

41. h"p://en.wikipedia.org/wiki/Cataract

42. h"p://en.wikipedia.org/wiki/File:Cataract_surgery.jpg
,A A
h"p://www.obihiroMmed.or.jp/blog/2007/12/postM304.html

43. (IOL)

44. ( )
IOL
"
"
QOV(Quality of vision)

45. [Holladay et al. , 1999]
[Franchini et al. , 2003]
[ , 2010]
4
#
#
#
#

46. 4
•
• 4

47. • 180
•
•
•

48. φ
θ
α
α
α
•
•
•

49. φ
θ
θ[deg.]
180
150
120
90
60
30
0 90 180 270 360
φ[deg.]

50. ( 30 )

51. ( 30 )

52. ( 30 )

53. ( 50 )

54. ( 50 )

55. ( 30 )

56. • CT
•
•
• NBI
h"p://sozai.rash.jp/medical/p/000154.htmlA
h"p://sozai.rash.jp/medical/p/000152.htmlA
h"p://medical.toykikaku.com/ / /

57. Designing Features and Classifiers for Colorectal Endoscopic
Images based on NBI Magnification Findings
NBI

58. ToruATamaki,AJunkiAYoshimuta,AMisatoAKawakami,ABisserARaytchev,AKazufumiAKaneda,AShigetoAYoshida,AYoshitoATakemura,AKeiichiAOnji,ARieAMiyaki,AShinjiA
Tanaka,AComputerMAidedAColorectalATumorAClassificaPonAinANBIAEndoscopyAUsingALocalAFeatures,AMedicalAImageAAnalysis,AAvailableAonlineA13ASeptemberA2012,A
ISSNA1361M8415,A10.1016/j.media.2012.08.003.A
A
YoshitoATakemura,AShigetoAYoshida,AShinjiATanaka,ARieAKawase,AKeiichiAOnji,AShiroAOka,AToruATamaki,ABisserARaytchev,AKazufumiAKaneda,AMasaharuAYoshihara,A
KazuakiAChayama,AComputerMaidedAsystemAforApredicPngAtheAhistologyAofAcolorectalAtumorsAbyAusingAnarrowMbandAimagingAmagnifyingAcolonoscopyA(withA
video),AGastrointesPnalAEndoscopy,AVolumeA75,AIssueA1,AJanuaryA2012,APagesA179M185,AISSNA0016M5107,A10.1016/j.gie.2011.08.051.A
KeiichiAOnji,AShigetoAYoshida,AShinjiATanaka,ARieAKawase,AYoshitoATakemura,AShiroAOka,AToruATamaki,ABisserARaytchev,AKazufumiAKaneda,AMasaharuAYoshihara,A
KazuakiAChayama,AQuanPtaPveAanalysisAofAcolorectalAlesionsAobservedAonAmagnifiedAendoscopyAimages,AJournalAofAGastroenterology,AVolumeA46,ANumberA12,A
1382M1390,A2011.A
A
,A ,ABisserARaytchev,A ,A ,A ,A :A NBI
,A PRMU2011M3,AVol.111,ANo.47,App.13M18,A ,A (2011A05).A
ToruATamaki,AJunkiAYoshimuta,ATakahishiATakeda,ABisserARaytchev,AKazufumiAKaneda,AShigetoAYoshida,AYoshitoATakemura,AShinjiATanaka:A"AAsystemAforA
ColorectalATumorAClassificaPonAinAMagnifyingAEndoscopicANBIAImages,"AProc.AofAACCV2010A;ATheA10thAAsianAConferenceAonAComputerAVision,AVol.2,App.987M998A
(2010A11),AQueenstown,ANewAZealand,ANovemberA8M12,A2010.A
,A ,A ,ABisserARaytchev,A ,A ,A ,A :A DenseASIFT NBI ,A
PRMU2010M73,AVol.110,ANo.187,App.129M134,A ,A (2010A09).A
A
YoshitoATakemura,AShigetoAYoshida,AShinjiATanaka,AKeiichiAOnji,AShiroAOka,AToruATamaki,AKazufumiAKaneda,AMasaharuAYoshihara,AKazuakiAChayama:A
"QuanPtaPveAanalysisAandAdevelopmentAofAaAcomputerMaidedAsystemAforAidenPficaPonAofAregularApitApa"ernsAofAcolorectalAlesions,"AGastrointesPnalA
Endoscopy,AVol.A72,ANo.A5,App.A1047M1051A(2010A11).A
MasashiAHIROTA,AToruATamaki,AKazuhumiAKaneda,AShigetoAYosida,AShinjiATanaka:A"FeatureAextracPonAfromAimagesAofAendoscopicAlargeAintesPne"AProc.AofA
FCV2008A;AtheA14thAKoreaMJapanAJointAWorkshopAonAFronPersAofAComputerAVision,App.94M99A(2008A01)A

59. • : 235,000 ( 21 )†
– 50,000A
fatali&es)of)colorectal)cancer)
40,000A
• : 42,434 ( )† 30,000A
– 20 1.7 20,000A
– 3 ( 1 : 2 : ) 10,000A
– 7 1
0A
'90A '91A '92A '93A '94A '95A '96A '97A '98A '99A '00A '01A '02A '03A '04A '05A '06A '07A '08A '09A
year)
†
• 5 : 20%
100AA
80AA
stage 1:
survival rate [%]
stage 2: 60AA
stage 3: 40AA
stage 4:
20AA
0AA
stageA1A stageA2A stageA3A stageA4A
stage 1 ( ) 100% 5 ‡
† http://www.mhlw.go.jp/toukei/saikin/
‡http://www.gunma-cc.jp/sarukihan/seizonritu/index.html

60. 8 10
h"p://www.mhlw.go.jp/toukei/saikin/hw/jinkou/geppo/nengai11/kekka03.html#k3_2

61. • CCD
•
I think this is a cancer…
100

62. 70 100

65. pit-pattern
• pit
– pit
–
pitMpa"ern [S.TanakaAetAal.,A‘06]
pit
pit
S pit
L pit
m sm
A
sm ) pitA
(
ASA ALA pit A
I pitA
sm
pit A
N A

66. NBI (NBI: Narrow-band Imaging)
• pit
–
–
NBI [H.Kanao et al., ‘09]
TypeAA
A
TypeAB pit
A
1 pit A
/
A
TypeAC 2 pit A
/ A
pit A
/ A
sm 3 (AVA) A
A

70. Our Project’s Goal
NBI
This Presentation’s Objective
NBI
<NBI >
Local binary patterns [Gross ‘08] : 90[%]
Vascularization features [Thomas ‘09]: 89.2[%]

71. texture analysis approach
YoshitoATakemura,AShigetoAYoshida,AShinjiATanaka,AKeiichiAOnji,AShiroAOka,AToruATamaki,AKazufumiAKaneda,AMasaharuAYoshihara,AKazuakiAChayama:A
"QuanPtaPveAanalysisAandAdevelopmentAofAaAcomputerMaidedAsystemAforAidenPficaPonAofAregularApitApa"ernsAofAcolorectalAlesions,"AGastrointesPnalA
Endoscopy,AVol.A72,ANo.A5,App.A1047M1051A(2010A11).A

72. Outline of Our Approach
+ Bag-of-features
Learning
Type A Type B Type C3
Test image
12, 55, 63, … 12, 55, 63, … 12, 55, 63, …
32, 20, 40, …
73, …
32, 20, 40, …
73, … 32, 20, 40, …
73, … 79, 5, 21, 19, 84, 99, 40, , 121
79, 5,21, 25,
87,27, 64, … …, 87 79, 5,21, 47,
87,66, 95, … …, 85 87,65, 33, … …, 101
…
67,49, 0, 87, …
11,6, …
36, 67,49, 0, 87, …
11,6, 82, 3, …, 124
36, … 67,49, 0, 87, …
11,6, …
36, 5, 26, 91, , 150
93, 41, 75, , 8
… 11, 52, 51, 32, , 89
…
…
…
…
…
…
Description of Local features Clusterin
g
Vector quantization
Vector quantization
Feature space
Type A Type B Type C3
Classifier
Histogram
Classification result

73. : gridSIFT
• Scale Invariant Feature Transform (SIFT) [Lowe, ‘99]
– 128
– DoG 90[%]
DoG
• grid sampling SIFT (gridSIFT)
–
scale size
– SIFT
grid sampling
grid space

74. : Support Vector Machine (SVM)
2
2
1 1 w 2
max
1 w
2 subject to yi w⋅ φ(x i ) ≥1
w w
• €
2
– Radial basis function (RBF) kRBF (u, v) = exp(−γ ⋅ u − v )
– linear klinear (u, v) = u" ⋅ v
– χ2 & γ (u − v )2 #
k χ 2 (u, v) = exp$ − ⋅
$ 2 u+v !
!
% "
• : One-Versus-One

75. •
• Type
• : 100×300 900×800[pix.]
• 2
< >
Type A:
Type B:
Type C3:

76. • Bag-of-features’s Approach
– : gridSIFT
– : k-means
– : SVM
• Dataset
– 908 NBI images (Type A: 359, Type B: 462, Type C3: 87)
• : 10-fold Cross Validation
– 8 900
– # of visual-words: 3×22, 3×23, …, 3×213

77. • SVM:
100AA
95AA
90AA
Correct)Rate)[%]
85AA
$ : χ2 > linear > RBF 80AA
75AA
1[%] 70AA
65AA
RBFMkernelA
linearMkernelA
χ2MkernelA
60AA
10A 100A 1000A 10000A 100000A
#)of)visual7words)[7]
100AA
• gridSIFT: 95AA
90AA
Correct)Rate)[%]
85AA
$ : > 0.005 80AA
75AA
… 70AA ContrastThreshold=0.005A
ContrastThreshold=0A
65AA
60AA
10A 100A 1000A 10000A 100000A
#)of)visual7words)[7]
100AA
• gridSIFT:
95AA
90AA
Correct)Rate)[%]
85AA
$ : 5[pix.] > 10[pix.] > 15[pix.]
80AA
75AA
gridAspace=5[pix.]A
70AA
gridAspace=10[pix.]A
65AA gridAspace=15[pix.]A
60AA
10A 100A 1000A 10000A 100000A
#)of)visual7words)[7]
• grisSIFT:
100AA
95AA
90AA
Correct)Rate)[%]
$ : (5, 7) [pix.] > other combination 85AA scale=3A
scale=5A
scale=7A
80AA scale=9A
scale=12A
scale=5,7A
75AA scale=5,9A
scale=5,12A
scale=7,9A
70AA
scale=7,12A
scale=9,12A
scale=5,7,12A
65AA
scale=5,7,9A
scale=5,9,12A
scale=7,9,12A
60AA
10A 100A 1000A 10000A 100000A
#)of)visual7words)[7]

78. • Bag-of-features’s Approach
– : gridSIFT (threshold: 0, grid space: 5[pix.], scale size: 5, 7[pix.])
– : k-means
– : SVM (linear kernel, -3 log2C 19)
• Dataset
– 908 NBI images (Type A: 359, Type B: 462, Type C3: 87)
• : 10-fold Cross Validation
– 8 900
– # of visual-words: 3×22, 3×23, …, 3×213

79. Results <10-fold Cross Validation>
100AA
Correct)Rate) Recall)Rate
100AA
96.00% 90AA
80AA
95AA
70AA
Recall)Rate)[%]
60AA
50AA
90AA
40AA
30AA TypeAAA
85AA 20AA TypeABA
Correct)Rate)[%]
10AA TypeAC3A
0AA
80AA 10A 100A 1000A 10000A 100000A
#)of)visual7words)[7]
100AA Precision)Rate
75AA 90AA
80AA
70AA
Precision)Rate)[%]
70AA
60AA
50AA
65AA 40AA
30AA TypeAAA
20AA TypeABA
60AA 10AA
TypeAC3A
10A 100A 1000A 10000A 100000A 0AA
10A 100A 1000A 10000A 100000A
#)of)visual7words)[7]
#)of)visual7words)[7]

80. • Bag-of-features’s Approach
– : gridSIFT (threshold: 0, grid space: 5[pix.], scale size: 5, 7[pix.])
– : k-means
– : SVM (linear kernel)
• Dataset
– 1412 NBI images: 908 training images (Type A: 359, Type B: 462, Type C3: 87)
504 test images (Type A: 156, Type B: 294, Type C3: 54)
• : Holdout Testing
–
– # of visual-words: 3×22, 3×23, …, 3×213

81. Results <Holdout Testing>
100AA Correct)Rate) 100AA Recall)Rate
90AA
92.86% 80AA
95AA
70AA
Recall)Rate)[%]
60AA
90AA 50AA
40AA
30AA
TypeAAA
85AA 20AA
Correct)Rate)[%]
TypeABA
10AA TypeAC3A
0AA
80AA 10A 100A 1000A 10000A 100000A
#)of)visual7words)[7]
75AA 100AA Precision)Rate
90AA
80AA
70AA 70AA
Precision)Rate)[%]
60AA
50AA
65AA
40AA
30AA
TypeAAA
20AA TypeABA
60AA
10A 100A 1000A 10000A 100000A 10AA TypeAC3A
0AA
10A 100A 1000A 10000A 100000A
#)of)visual7words)[7] #)of)visual7words)[7]

82. Conclusions
•
$ gridSIFT:
$ SVM:
96.00[%] (10-fold Cross Validation)
Future Works
• Type C3
$
•

83. Real-Time Recognition System
for NBI Video Endoscopy
,A ,A ,ARaytchevABisser,A ,A ,A ,A :A NBI
,A 17 SSII2011,App.IS1M09M1MIS1M09M7,A ,A (2011A06).A

84. 120[pix.]
•
120[pix.]
22 6 … :14.7[fps]
91 87 …
•
• SIFT
A
Visual Word Histogram B
C3
C3
• : SVM A
B
• (A or B or C3)
• A, B, C3

85. • Dataset • Dataset
A B C3 A B C3
359 461 87 907 4 5 3 12
640*480 [pix.]
120*120 [pix.]
200 2400
• SIFT Dense SIFT(VLFeat) • Visual word 768
— : 5[pix.]
— : 5, 7[pix.]
• SVM(LibSVM)
— : Linear

86. probability
0
1
A
B
C3
time

87. MRF
NBI
Temporal labeling NBI Videoendoscopy Using MRF
,A ,A ,ABisserARaytchev,A ,A ,A ,A ,A ,A NBI ,A
18 SSII2012,App.IS1M09M1MIS1M09M5,A ,A (2012A06).A

88. [ ]
•
1
Probability
Type A
0.5 Type B
Type C3
0
0 20 40 60 80 100 120 140 160 180 200
フレーム番号
•
Type A
Type B
Type C3

89. Goal
%
This Presentation’s Objective
SVM MRF

90. MRF
# & # &
f ( x y ) ∝ exp % ∑ A ( xi , yi ) ( ⋅ exp % ∑ I ( xi , x j ) (
% (
$ i ' $ j∈N i '
B B C3 C3 B
x1 ………… x50 ………… x100 ………… x150 ………… x200
0 50 100 150 200 i
…… …… …… ……
y1 y50 y100 y150 y200
x:
y: SVM

91. MRF
# & # &
f ( x y ) ∝ exp % ∑ A ( xi , yi ) ( ⋅ exp % ∑ I ( xi , x j ) (
% (
$ i ' $ j∈N i '
SVM
1
Probability
Type A
0.5 Type B
Type C3
0
0 20 40 60 80 100 120 140 160 180 200
フレーム番号
P(x50=A|y50) = 0.004
P(x50=B|y50) = 0.99 exp ( A ( xi , yi )) = P ( xi yi )
P(x50=C3|y50) = 0.006
A ( xi , yi ) = log P ( xi yi )

92. MRF
# & # &
f ( x y ) ∝ exp % ∑ A ( xi , yi ) ( ⋅ exp % ∑ I ( xi , x j ) (
% (
$ i ' $ j∈N i '
• • C3
ー Type ー Type C3
ー Type C3
y1 yi−1 yi yi+1 yn y1 yi−1 yi yi+1 yn
x1 xi−1 xi xi+1 xn x1 xi−1 xi xi+1 xn

93. MRF
# & # &
f ( x y ) ∝ exp % ∑ A ( xi , yi ) ( ⋅ exp % ∑ I ( xi , x j ) (
% (
$ i ' $ j∈N i '
Label B C3 ? B B xi−1 xi
Time A
i − 2 i −1 i i +1 i + 2 B
C3 C3
Label B B B B B

94. MRF
# & # & &
f ( x y ) ∝ exp % ∑ A ( xi , yi ) ( exp % ∑ I I xixhx jxi( x j ) (
% ∑ ( ( , , )( ( ,
$ i ' $ h, j∈Ni
j∈N i ' '
• • C3
ー Type ー Type C3
ー Type C3
y1 yi−1 yi yi+1 yn y1 yi−1 yi yi+1 yn
x1 xi−1 xi xi+1 xn x1 xi−1 xi xi+1 xn

95. MRF
# & # &
f ( x y ) ∝ exp % ∑ A ( xi , yi ) ( ⋅ exp % ∑ I ( xh , xi , x j ) (
% (
$ i ' $ h, j∈N i '
C3
Label B C3 ? B B xi−1 xi xi+1
Time A A
i − 2 i −1 i i +1 i + 2 B
C3 C3
Label B C3 C3 B B

96. MRF
# & # &
f ( x y ) ∝ exp % ∑ A ( xi , yi ) ( ⋅ exp % ∑ I ( xi , x j ) (
% (
$ i ' $ j∈N i '
(MAP) x
%
• • C3
ー Type ー Type C3
ー Type C3
y1 yi−1 yi (DP)
yi+1 yn y1 yi−1 yi yi+1 yn
x1 xi−1 xi xi+1 xn x1 xi−1 xi xi+1 xn

97. [ ]
• 907
(Type A: 359, Type B: 462, Type C3: 87)
•
• Type
• 2
•
•
200
• 4
(Type A: 2 Type B: 2 )

98. Type B (original)
Type A Type B
1 1
0 20 40 60 80 100 120 140 160 180 200
0.5 0.5 frame number
Type B (DP_0.8)
Type B (Gibbs_p4=0.6)
0 0
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
A B
B A
C 0 20 40 60 80 100 120 140 160 180 200
C
frame number
Type A_1 (original) Type B (original)
Type B (DP_0.9)
Type B (Gibbs_p4=0.7)
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
frame number frame number
Type A_1 (DP_0.99) Type B (DP_0.8)
Type B (DP_0.99)
Type B (Gibbs_p4=0.8)
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
frame number frame number
Type A_1 (Gibbs_p4=0.9) Type BB (DP_0.9)
Type (DP_0.999)
Type B (Gibbs_p4=0.9)
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
frame number frame number
Type B (DP_0.99)
Type A Type B Type C3
0 20 40 60 80 100 120 140 160 180 200

99. Type B (original)
Type A Type B
1 1
0 20 40 60 80 100 120 140 160 180 200
0.5 0.5 frame number
Type B (DP_0.8)
Type B (Gibbs_p4=0.6)
0 0
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
A B
B A
C3
C3
0 20 40 60 80 100 120 140 160 180 200
C C
frame number
MAP
Type A_1 (original) Type B (original)
Type B (DP_0.9)
Type B (Gibbs_p4=0.7)
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
frame number frame number
Type A_1 (DP_0.99) ( )Type B (Gibbs_p4=0.8)
Type B (DP_0.8)
Type B (DP_0.99)
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
frame number frame number
Type A_1 (Gibbs_p4=0.9) (C3 )
Type BB (DP_0.9)
Type (DP_0.999)
Type B (Gibbs_p4=0.9)
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
frame number frame number
Type B (DP_0.99)
Type A Type B Type C3
0 20 40 60 80 100 120 140 160 180 200

100. A
B
C3
1A
Probability
TypeAAA
0.5A MRF
TypeABA
0A Type A Type B Type C3 TypeAC3A
0A 50A 100A 150A 200A

101. Type A Type B
Type A Type B
Type A Type B Type C3

102. Self-training
~ NBI ~
Self-training with unlabeled regions and
its application to recognition of colorectal NBI
endoscopic images
,A ,A ,A ,A ,A ,A ,A ,A ,A
SelfMtrainingA NBIA ,A ,A PRMU2012M11,AVol.112,A
No.37,App.57M62,A ,A (2012A05).A

103. MOTIVATION
•
!
! NBI
!
×
C3
×
×
NBI

104. ABSTRACT
Key Idea :
• Self-training
&
&
[Yoshimuta et al., ‘10]

105. Self-training
•
•
Accept POINT
1.
Reject
2.

106. labeled samples
•
• 100×300 900×800 [pix.]
•
Type)A Type)B Type)C3 Total
359 462 87 908
B C3
A

107. Unlabeled samples
• 10
• 30×30 250×250 [pix.]
•
–
–
•
Type)A Type)B Type)C3 Total
3590 4610 870 9070
* 10

108. Evaluation
# 10 hold out testing 10
# t
(1) + (5) + (9)
=
(1) + (2) + (3) + (4) + (5) + (6) + (7) + (8) + (9)
(9)
C3 =
(7) + (8) + (9)
Estimated Category
Type A Type B Type C3
True Type A (1) (2) (3)
Category Type B (4) (5) (6)
Type C3 (7) (8) (9)

109. result
#
&
& C3
#
&
&

110. Result
0.96A
p=0.013314
0.95A
Recogni&on)Rate)
0.94A
0.93A
0.92A
0.91A
0.9A
AlgorithmA1A AlgorithmA2A AlgorithmA3A

111. • CT
•
•
• NBI
h"p://sozai.rash.jp/medical/p/000154.htmlA
h"p://sozai.rash.jp/medical/p/000152.htmlA
h"p://medical.toykikaku.com/ / /

112. h"p://kids.wanpug.com/top_person.html