2. Semi-
supervised Proposed Experimental
Introduction Algorithm Setting Result Conclusion
Learning
ENDOSCOPIC
DIAGNOSIS
CCD
I think this is a cancer…
100
4. Semi-
supervised Proposed Experimental
Introduction Algorithm Setting Result Conclusion
Learning
NBI CLASSIFICATION
! Stehle et al., ’09 :
! Gross et al., ’09 :
! Tamaki ACCV2010, PRMU2011 : Bag-of-Visual
Words
NBI [H. Kanao et al., ‘09]
hyperplasia (HP) Type A
Stehle et al.
tubular adenoma(TA) Gross et al. Type B
PRMU2011
M~SM-s
SM-s Type C3
10. Proposed Experimental
Introduction Self-training Algorithm Setting Result Conclusion
ALGORITHM 1
Self-training (estimate probability)
Labeled samples L Unlabeled samples U
Estimate label EL j
Classifier f A B A B A C3
Estimate probability EPj
0.5 0.7 0.9 0.9 0.9 0.6
EPj ≥ TH = 0.9
EL
f B A B
EP
0.9 0.8 0.5
EPj ≥ TH = 0.9
11. Semi-
supervised Proposed Experimental
Introduction Algorithm Setting Result Conclusion
Learning
ORIGINAL LABEL
CONSTRAINT
• Type A Type B, Type C3
• Type B Type C3
13. Proposed Experimental
Introduction Self-training Algorithm Setting Result Conclusion
ALGORITHM 2
Self-training (estimate probability & estimate label)
l +u
Labeled samples L Unlabeled samples U Original labels{ y j } j =l +1
B A A B B C3
Estimate label EL j
Classifier f A B A B A C3
Estimate probability EPj
0.5 0.7 0.9 0.9 0.9 0.6
EPj ≥ TH = 0.9 y j = EL j
B A B C3
EL
f B A B B
EP
0.9 0.8 0.8 0.5
EPj ≥ TH = 0.9 y j = EL j
15. Proposed Experimental
Introduction Self-training Algorithm Setting Result Conclusion
ALGORITHM 3
Self-training
Labeled samples L
l
Labels { yi }i =1 Unlabeled samples U
d( x i , x j ) yi B A A C3 A A
128 min d( xi , x j ) 0.9 0.7 1.9 1.5 2.9 2.6
d( x i , x j ) = ∑ (x id − x jd ) 2 min d( xi , x j ) < 1.5
d =1
Classifier f
17. Semi-
supervised Proposed Experimental
Introduction Algorithm Setting Result Conclusion
Learning
LABELED SAMPLES
100×300 900×800 [pix.]
Type A Type B Type C3 Total
359 462 87 908
B C3
A
18. Semi-
supervised Proposed Experimental
Introduction Algorithm Setting Result Conclusion
Learning
UNLABELED SAMPLES
10
30×30 250×250 [pix.]
•
•
Type A Type B Type C3 Total
3590 4610 870 9070
* 10
19. Semi-
supervised Proposed Experimental
Introduction Algorithm Setting Result Conclusion
Learning
UNLABELED SAMPLES
10
30×30 250×250 [pix.]
•
•
Type A Type B Type C3 Total
3590 4610 870 9070
* 10