[DL輪読会]A Probabilistic U-Net for Segmentation of Ambiguous Images

DEEP LEARNING JP A Probabilistic U-Net for Segmentation of Ambiguous Images [DL Papers] Tomoki Tanimura, Keio University 1 http://deeplearning.jp/

• http://deeplearning.jp/

RNS § .0 § P 0 8: § § 81 , : 1 2 I , 12 MCDAG MC K

§ § § § § §

§ § § § § §

§ § §

§ § § § § §

N , ,- , § § § § M C : , F IS R ,- , , , R

§ § § § § §

G § § § § R § § T R - C

C § § - 1 G = § - I

= § § G ( ) Graders

§ § § § § §

+ § -

+ § U-Net -

§ + Segmentation

§ -+ -

§ + -

§ C - + - + AE

C § + ( ) § Segmentation Variant6 § E 3 § U-Net z6 § ) 1 6 - A

-

- ①Prior Net Posterior

- ②segmentation map

ν ω § Posterior (Prior ) z θ ψ X S - Y : Loss Loss

§ § § § § §

§ § C [ C R B § § ( - § § ) ) § D ) 4 § § LD T E G I .

2 2 - - - - - - -

§

§ § -

§ §

] § 2 . § § 2 2 2 . s V § d epS tI E E h P Irm v A xn fa gl [ .2 I o ci

U § (2 2 21 2 = § Y a § : (, § § 1 = S : 2 21 T : ) = d 21 21 E S dP GI d cD (,

T § (2 2 § § § § 21 c Y : (, 1 = Ud 2 = : a ) : = 2 21 E 21 a S 21 Pred S (, G Ud GT P eD 2I Pred S 2I GT S

GT Pred 2 Pred 2 GT

§ GT § Pred GT Pred 2 Pred 2 GT

§ § 2 ) § G 2 G ) G () 1 § GT2Pred 2 Pred 2 GT

§ = §G 2 = § § G 4

§ § d 1/: 2/ / § e 1/: af G/:: / 7 : § /7:7:5 ./972/ 7 : § 4 § § § b 3 2 § V § § § § § /98C :C 1/ C 1/ V 72 c gGT 72 C /98 C 7 6 / / 797 : C 76/ / 797 4 76 5 / 7 :C 5 /7 : C 76 /2C /2 C 7 6 / 797 , 4

g § § § n D § § d § m 4 3 § S e D 3 D T i 3 § a § o D 3 G 2 3

( § )

a § , ,6 14 § PD N 8 a - bS U § § § e N ) - ( 4,1 4 d

e § ) cbI § N § ) ( § ( § a )C( U - P :I T 1 T U ) T G I ( d C (1

U § § § D § - § EN EN U N U

I- E § H § § M IA § I- 2 V A

N § § § N § G - -

() ( § () ( 3 § ( 2 C

N § iT r § - § S o Gt e n V Ug § iT r a Prior V Prior m P n Posterior U-Net fix GT

N § G A - - - A § UT - V - A E e S a

§ § § § § §

§ §s N Gi n E - §t + o §r - U n e ya ec e - + S N P g pm N - r - + U TV S TV AC + - C