U-Net

๋ถ„์•ผ
Network Architecture
๋ฆฌ๋ทฐ ๋‚ ์งœ
2020/08/17
๋ณธ ํฌ์ŠคํŠธ๋Š” ์ œ๊ฐ€ ํœด๋จผ์Šค์ผ€์ดํ”„ ๊ธฐ์ˆ  ๋ธ”๋กœ๊ทธ์— ๋จผ์ € ์ž‘์„ฑํ•˜๊ณ  ์˜ฎ๊ธด ํฌ์ŠคํŠธ์ž…๋‹ˆ๋‹ค.
๋ณธ ํฌ์ŠคํŠธ์—์„œ๋Š” biomedical segmentation ๋ถ„์•ผ์—์„œ ์ธ์šฉ ์ˆ˜ 16000+์„ ์œก๋ฐ•ํ•˜๋Š” ๋…ผ๋ฌธ์— ๋Œ€ํ•ด์„œ ๋ฆฌ๋ทฐํ•˜๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์˜ ์ œ๋ชฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
โ€œU-Net: Convolutional Networks For Biomedical Image Segmentationโ€
๋…ผ๋ฌธ์— ๋Œ€ํ•œ ๋‚ด์šฉ์„ ์ง์ ‘ ๋ณด์‹œ๊ณ  ์‹ถ์œผ์‹  ๋ถ„์€ย ์ด๊ณณ์„ ์ฐธ๊ณ ํ•˜์‹œ๋ฉด ์ข‹์Šต๋‹ˆ๋‹ค.
Network Architecture of U-Net

Objective

๋…ผ๋ฌธ์—์„œ ๋ชฉ์ ์œผ๋กœ ํ•˜๋Š” ๋ฐ”๋Š” biomedical image segmentation์ž…๋‹ˆ๋‹ค. ์ž์„ธํžˆ ์„ค๋ช…ํ•˜์ž๋ฉด, ์ „์ž ํ˜„๋ฏธ๊ฒฝ ๋“ฑ์œผ๋กœ ๊ด€์ฐฐํ•œ ์ด๋ฏธ์ง€์—์„œ ํŠน์ • ์„ธํฌ(ex. ์‹ ๊ฒฝ์„ธํฌ, ์•”์„ธํฌ)๋ฅผ ๋ฐฐ๊ฒฝ์œผ๋กœ๋ถ€ํ„ฐ ๋ถ„๋ฅ˜ํ•˜๋Š” ์ž‘์—…์ž…๋‹ˆ๋‹ค. ๋ณธ๋ž˜ segmentation์ด๋ผ ํ•˜๋ฉด object detection์„ ํฌํ•จํ•˜๋Š” ๊ฐœ๋…์œผ๋กœย ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ๊ฐ์ฒด๋“ค์„ ๋ถ„๋ฅ˜ํ•ด๋‚ด๊ณ  ์ฐพ์•„๋‚ด๋Š” ๊ฒƒ์ด์ง€๋งŒ, ๋…ผ๋ฌธ์—์„œ ์ดˆ์ ์„ ๋งž์ถ”๊ณ ์ž ํ•˜๋Š” ๋ฐ”๋Š” ์ด๋ฏธ์ง€ ์†์—์„œ ๋‹ค์–‘ํ•œ ์„ธํฌ๋ฅผ ๋ถ„๋ฅ˜ํ•ด ๋‚ด๋Š” ๊ฒƒ์ด ์•„๋‹Œ ๋‹จ์ˆœํžˆย ํ•œ ์ข…๋ฅ˜์˜ ๊ฐ์ฒด๋“ค์„ ๋ฐฐ๊ฒฝ๊ณผ ์™„์ „ํžˆ ๋ถ„๋ฆฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๊ตฌ์กฐ์  ๊ด€์ ์˜ ๋ชฉ์ ์€ U-Net architecture ์˜ ๋งˆ์ง€๋ง‰ layer์˜ output์œผ๋กœ ์‚ฐ์ถœ๋˜๋Š”ย segmenation map์„ ํ•™์Šต์„ ํ†ตํ•ด ์–ป์–ด๋‚ธ๋‹ค๊ณ  ๋ณด์‹œ๋ฉด ๋ฉ๋‹ˆ๋‹ค.

Comparison with Prior Study

๋…ผ๋ฌธ์—์„œ ์ฃผ๋กœ ๋น„๊ต์˜ ๋ชฉ์ ์œผ๋กœ ๊บผ๋‚ด๋Š” ๋…ผ๋ฌธ์ด Ciresan et al.(์ดํ•˜ [1])์ž…๋‹ˆ๋‹ค. [1]์˜ architecture๋Š” EM segmentation challenge at ISBI 2012์—์„œ ์šฐ์Šน์„ ์ฐจ์ง€ํ–ˆ์„๋งŒํผ ์„ฑ๊ณต์ ์ด์—ˆ์ง€๋งŒ, ํฌ๊ฒŒ ๋‘ ๊ฐ€์ง€์˜ ๋ฌธ์ œ์ ์ด ์กด์žฌํ–ˆ์Šต๋‹ˆ๋‹ค.
์ฒซ ๋ฒˆ์งธ๋Š” localization(์ด๋ฏธ์ง€์˜ ์„ธ๋ถ€ ํŠน์„ฑ ๋ฐ˜์˜)์„ ์œ„ํ•ด์„œ image๋ฅผ patch๋‹จ์œ„๋กœ ๋ถ„๋ฆฌํ•˜์—ฌ window-slide ํ˜•ํƒœ๋กœ ํ•™์Šต์‹œ์ผœ์„œ ๋ฐ์ดํ„ฐ์˜ ์ค‘๋ณต์ด ๋งค์šฐ ๋งŽ์œผ๋ฉฐ, ํ•™์Šต ์‹œ๊ฐ„์ด ๊ธธ์—ˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค.
Window-slide
Window-slide๋Š” ์œ„ ๊ทธ๋ฆผ์ฒ˜๋Ÿผ image๋ฅผ patch๋กœ ๋‚˜๋ˆ„์–ด ํ•™์Šต์‹œํ‚ฌ ๋•Œ ๊ณตํ†ต๋œ ๋ถ€๋ถ„์ด ๊ต‰์žฅํžˆ ๋งŽ๊ณ , ์ด์— ๋”ฐ๋ผ patch ๊ฐœ์ˆ˜๋„ ๋งŽ์•„์ง‘๋‹ˆ๋‹ค. ํšจ๊ณผ์ ์ธ localization์„ ์œ„ํ•ด์„œ๋Š” ์ข‹์€ ๋ฐฉ๋ฒ•์ผ ์ˆ˜ ์žˆ์œผ๋‚˜, ํ•™์Šต ์†๋„ ์ธก๋ฉด์—์„œ ์ข‹์€ ๋ฐฉ๋ฒ•์ด ์•„๋‹™๋‹ˆ๋‹ค.
๋‘ ๋ฒˆ์งธ๋Š” localization๊ณผ context๊ฐ„์˜ trade-off๊ฐ€ ์กด์žฌํ•˜์—ฌ ๋‘ ํŠน์„ฑ์„ ๋ชจ๋‘ ์ฑ™๊ธฐ๊ธฐ ์–ด๋ ค์› ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค.
Patch size๊ฐ€ ํฌ๋ฉด feature data๊ฐ€ context(์ด๋ฏธ์ง€์˜ ๊ฐœ๋žต์ ์ธ ํŒจํ„ด ๋ฐ˜์˜)์— ๋Œ€ํ•œ ๋‚ด์šฉ์€ ํฌํ•จํ•˜๊ธฐ ์šฉ์ดํ•˜์ง€๋งŒ localization์— ๋Œ€ํ•œ ๋‚ด์šฉ์„ ํฌํ•จํ•˜๊ธฐ๋Š” ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์—, patch size๊ฐ€ ์ž‘์œผ๋ฉด localization์— ๋Œ€ํ•œ ๋‚ด์šฉ์€ ํฌํ•จํ•˜๊ธฐ ์šฉ์ดํ•˜์ง€๋งŒ context์— ๋Œ€ํ•œ ๋‚ด์šฉ์€ ํฌํ•จํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค.

Methods

์•ž์„œ ์ œ์‹œํ•œ ์„ ํ–‰์—ฐ๊ตฌ์˜ ๋ฌธ์ œ์ ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ, ๋…ผ๋ฌธ์—์„œ๋Š” ํฌ๊ฒŒ ๋‘ ๊ฐ€์ง€์˜ ๊ฐœ์„ ๋ฐฉ์•ˆ๊ณผ, ๋ชฉ์ ์„ ํšจ๊ณผ์ ์œผ๋กœ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์„ค๊ณ„๋ฅผ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
์ฒซ ๋ฒˆ์งธ ๋ฌธ์ œ์ ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ ๋…ผ๋ฌธ์—์„œ๋Š” Overlap-Tile Strategy๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.

Overlap-Tile Strategy

Overlap-Tile Strategy๋Š” ์œ„ ๊ทธ๋ฆผ์ฒ˜๋Ÿผ ํŒŒ๋ž€์ƒ‰ ๋ฒ”์œ„์˜ input image patch๋ฅผ ๋„ฃ์—ˆ์„ ๋•Œ ๋…ธ๋ž€์ƒ‰ ๋ฒ”์œ„์˜ segmented image๊ฐ€ ๋‚˜์˜จ๊ฒŒ ๋˜๋Š” ์ƒํ™ฉ์—์„œ, output segmented image๊ฐ€ ๊ฒน์น˜๋Š” ๋ฒ”์œ„ ์—†์ด ๋‚˜์˜ค๋„๋ก input image patch์˜ ์„ค์ • ๋ฒ”์œ„๋ฅผ ์กฐ์ ˆํ•˜๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ์šฐ์ธก์˜ ๋‘ ๋…ธ๋ž€์ƒ‰ ๋ฒ”์œ„๊ฐ€ ๋“ฑ์žฅํ•˜๋Š” output segmented image์ผ ๋•Œ ์„ค์ •ํ•ด์•ผ ํ•  ๋‘ input image patch์˜ ๋ฒ”์œ„๋Š” ์šฐ์ธก์˜ ๋‘ ํŒŒ๋ž€์ƒ‰ ๋ฒ”์œ„์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค. Overlap์ด๋ผ๋Š” ๋‹จ์–ด๋Š” padding์ด ์กด์žฌํ•˜์ง€ ์•Š๋Š” ๋…ผ๋ฌธ์˜ architecture ํŠน์„ฑ์ƒ input์— ๋น„ํ•ด output์˜ size๊ฐ€ ์ค„์–ด๋“ค๊ธฐ ๋•Œ๋ฌธ์— ๋‚˜ํƒ€๋‚˜๋Š” input ์„ค์ •์˜ ๊ฒน์น˜๋Š” ํŠน์„ฑ์„ ์ผ์ปซ๋Š” ๋ง์ž…๋‹ˆ๋‹ค.
์ด์™€ ๊ฐ™์ด patch size๋ฅผ ์„ค์ •ํ•˜์—ฌ ๊ณผ๋„ํ•˜๊ฒŒ ๋งŽ์€ patch ๊ฐœ์ˆ˜๋ฅผ ์ค„์ด๋Š” ์ „๋žต์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์— ๋”๋ถˆ์–ด ์ตœ์ข…์ ์œผ๋กœ ๋นˆ ๋ถ€๋ถ„์ด ๋˜์–ด์ง€๋Š” ๊ฐ€์žฅ์ž๋ฆฌ ๋ถ€๋ถ„์€ mirroring extrapolation(๊ฐ€์žฅ์ž๋ฆฌ๋ฅผ ๊ฑฐ์šธ ๋Œ€์นญ)์„ ์ด์šฉํ•ด์„œ ์ฑ„์šฐ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.
U-Net Architecture๋‘ ๋ฒˆ์งธ ๋ฌธ์ œ์ ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ ๋…ผ๋ฌธ์—์„œ๋Š” U-Net Architecture๋ฅผ ๊ตฌํ˜„ํ–ˆ์Šต๋‹ˆ๋‹ค. U-Net Architecture๊ฐ€ ๊ธฐ์กด๊ณผ ๊ฐ€์žฅ ๋‹ค๋ฅธ ํŠน์ดํ•œ ์ ์€ Contracting Path, Exansive Path, ๊ทธ๋ฆฌ๊ณ  Contracting Path๋กœ๋ถ€ํ„ฐ Expansive Path๋กœ ์ฃผ์–ด์ง€๋Š” Skip Connection์ด ์กด์žฌํ•œ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค.

Detail U-Net Architecture

Contracting Path๋Š” ์ผ๋ฐ˜์ ์ธ deep learning architecture์—์„œ ๋ณด์ด๋Š” ๋ฐ”์™€ ๊ฐ™์ด image๋กœ๋ถ€ํ„ฐ feature๋ฅผ extractingํ•˜๋Š” ๋ถ€๋ถ„์ž…๋‹ˆ๋‹ค. Expansive Path๋Š” segmentation map(๊ฐ pixel์— ๋Œ€ํ•œ binary classification map)์„ ๋‹ค์‹œ ๋งŒ๋“ค์–ด๋‚ด๊ธฐ ์œ„ํ•ด์„œ ๋‹ค์‹œ upsampling ํ•˜๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค.
์—ฌ๊ธฐ๊นŒ์ง€๋Š” deep learning์˜ ๊ธฐ๋ณธ์ ์ธ ์„ค๊ณ„์ธ feature extraction๊ณผ ๋ชฉ์ ์„ฑ์— ๋งž๊ฒŒ upsamplingํ•˜๋Š” ๊ณผ์ •์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ์ง€๋งŒ, U-Net์—์„œ๋Š” ํŠน๋ณ„ํ•˜๊ฒŒ ์ž‘์€ patch size์—์„œ๋„ feature๊ฐ€ context์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ํฌํ•จํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜๊ธฐ ์œ„ํ•ด์„œ Skip Connection์ด๋ผ๋Š” ๊ฒƒ์„ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
Contracting Path์— ์กด์žฌํ•˜๋Š” ๊ฐ layer๋‹จ๊ณ„์— ์กด์žฌํ•˜๋Š” ์ผ๋ถ€ feature๋ฅผ ์ž˜๋ผ์„œ Expansive Path์˜ feature์— concatenateํ•ด์ฃผ์–ด ๊ฒฐ๊ณผ์ ์œผ๋กœ upsampling ๋‹จ๊ณ„์—์„œ ๋Š˜์–ด๋‚œ channel ์ˆ˜๊ฐ€ context์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ๊ฐ€์งˆ ์ˆ˜ ์žˆ๊ฒŒ ์„ค๊ณ„๋ฅผ ์ง„ํ–‰ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
์ด๋ ‡๊ฒŒ ์ „๋ฐ˜์ ์œผ๋กœ U ํ˜•ํƒœ๋ฅผ ๊ทธ๋ฆฌ๋Š” architecture๋ฅผ ๊ตฌํ˜„ํ•˜์—ฌ U-Net์ด๋ผ๋Š” ์ด๋ฆ„์„ ๊ฐ€์ง€๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.
Touching Object Classification๋…ผ๋ฌธ์—์„œ ์ง๋ฉดํ–ˆ๋˜ ๋ฌธ์ œ ์ค‘ ํ•˜๋‚˜๋Š” ๊ฒน์ณ์ ธ ์žˆ๋Š” ๊ฐ์ฒด๋“ค์˜ ํšจ๊ณผ์ ์ธ ๋ถ„๋ฅ˜์˜€์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋…ผ๋ฌธ์—์„œ๋Š” ํ•™์Šต์„ ์œ„ํ•œ Cross Entropy๋ฅผ ์ •์˜ํ•  ๋•Œ ๊ฐ€์žฅ์ž๋ฆฌ์™€ ๊ฐ€๊นŒ์šด pixel์— ๋”์šฑ ๊ฐ€์ค‘์น˜๋ฅผ ๋‘์–ด ํ•™์Šต์„ ์ง„ํ–‰์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ๋’ค์—์„œ ์ด์–ด์„œ ์ง„ํ–‰ํ•˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
Touching Object Classification

Training

๋…ผ๋ฌธ์—์„œ weight ํ•™์Šต์„ ์œ„ํ•œ ๊ธฐ์ค€์ด ๋˜๋Š” energy function์€ pixel-wise soft-max function๊ณผ cross-entropy loss function์œผ๋กœ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
๋จผ์ €ย pixel-wise soft max function์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
pk(X)=eak(X)โˆ‘kโ€ฒ=1Keak(X)p_k(X) = \frac{e^{a_k(X)}}{\sum_{k'=1}^K e^{a_k(X)}}
K๋Š” ์ „์ฒด label์˜ ์ˆ˜, k๋Š” ๊ทธ ์ค‘ ํŠน์ • label์„ ์ง€์ •ํ•˜๋Š” ๋ณ€์ˆ˜์ด๋ฉฐ, ๊ฒฐ๊ณผ์ ์œผ๋กœ ๋งˆ์ง€๋ง‰ output channel์˜ index๋ผ๊ณ  ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. a_k(X)๋Š” X์˜ ์œ„์น˜(pixel x,y)์— ์žˆ๋Š” activation์˜ ๊ฐ’์ž…๋‹ˆ๋‹ค.
์ด soft max function์˜ ์„ค๊ณ„์—์„œ ๊ธฐ๋Œ€ํ•˜๋Š” ๊ฒƒ์€ ํŠน์ • pixel์˜ ์œ„์น˜ X์— ๋Œ€ํ•ด ์˜ฌ๋ฐ”๋ฅธ k๋กœ ํ•™์Šต๋˜์—ˆ์„ ๋•Œ p_k(X)๊ฐ€ 1์— ๊ฐ€๊นŒ์šด ๊ฐ’์— ์ˆ˜๋ ดํ•˜๊ณ , ๋‹ค๋ฅธ ๊ฒฝ์šฐ์—” 0์— ๊ฐ€๊นŒ์šด ๊ฐ’์œผ๋กœ ์ˆ˜๋ ดํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋‹ค์Œ์œผ๋กœย cross-entropy loss function์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
E=โˆ‘Xโˆˆฮฉฯ‰(X)logโก(pl(X)(X))E = \sum_{X\in\Omega} \omega(X)\log(p_{l(X)}(X))
w(X)ํ•ญ์€ ๊ฐ€์ค‘์น˜์— ๋Œ€ํ•œ ํ•ญ๋ชฉ์ด๋ฉฐ l(X)๋Š” ์œ„์น˜ X์— ๋Œ€ํ•œ ground truth label์ž…๋‹ˆ๋‹ค. ์ „์ฒด function์˜ ํ˜•ํƒœ๋Š” ํ”ํžˆ ์•Œ๊ณ  ์žˆ๋Š” cross-entropy์˜ ํ˜•ํƒœ์™€ ๋งค์šฐ ์œ ์‚ฌํ•˜๋ฉฐ ๋‹ค๋ฅธ ์ ์ด ์žˆ๋‹ค๋ฉด ๊ฐ€์ค‘์น˜์— ๋Œ€ํ•œ ํ•ญ๋ชฉ์ž…๋‹ˆ๋‹ค. ๊ฐ ์œ„์น˜์— ๋Œ€ํ•œ ground truth channel์˜ soft max function์˜ logarithm์„ ํ•ฉ์‚ฐํ•œ ํ˜•ํƒœ๋กœ ๋ณผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, soft max function ๊ฒฐ๊ณผ๊ฐ€ ์ œ๋Œ€๋กœ ์˜ˆ์ธก๋  ์ˆ˜๋ก ์—๋„ˆ์ง€๊ฐ€ ์ปค์ง€๋Š” ํ˜•ํƒœ์ž„์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋”๋ถˆ์–ด ์•ž์„œ ์ƒ๋žตํ•œย ๊ฐ€์ค‘์น˜์— ๋Œ€ํ•œ ์„ค๋ช…์„ ๊ฐ„๋‹จํžˆ ๋“œ๋ฆฌ์ž๋ฉด, ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ˆ˜์‹์œผ๋กœ ์ •์˜ํ–ˆ์Šต๋‹ˆ๋‹ค.
ฯ‰(X)=ฯ‰c(X)+ฯ‰0eโˆ’(d1(X)+d2(X))22ฯƒ2\omega(X) = \omega_c(X)+\omega_0e^{-\frac{(d_1(X)+d_2(X))^2}{2\sigma^2}}
w_c(X)๋Š” ๊ฐ training data ๋ณ„๋กœ class frequency(labeling๋œ pixel์˜ ๋น„์œจ๋“ค)์ด ๋‹ค๋ฅธ ๊ฒƒ์„ ์กฐ์ ˆํ•ด ์ฃผ๋Š” weight map์ž…๋‹ˆ๋‹ค. d_1(X)๋Š” ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด cell๊ณผ border๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ์ด๊ณ  d_2(X)๋Š” ๋‘ ๋ฒˆ์งธ๋กœ ๊ฐ€๊นŒ์šด cell๊ณผ border๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์ฃผ๋ชฉํ•  ์ ์€ border๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ๊ฐ€ ๊ฐ€๊นŒ์šด ์œ„์น˜์— ์žˆ๋Š” pixel๋“ค์ด ๊ธฐ์—ฌํ•˜๋Š” ์—๋„ˆ์ง€์— ๋” ๊ฐ€์ค‘์น˜๊ฐ€ ๋†’๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์—์„œ๋Š” ์ด ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด์„œ ์ธ์ ‘ํ•œ ๊ฐ์ฒด๋“ค์˜ ๊ฒฝ๊ณ„๋ฅผ ํ™•์‹คํžˆ ๊ตฌ๋ณ„ํ•˜๊ธฐ ์œ„ํ•œ ๊ณผ์ œ์˜€๋˜ touching object classification ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋ ค๊ณ  ํ–ˆ์Šต๋‹ˆ๋‹ค.
๋งˆ์ง€๋ง‰์œผ๋กœ, ๋…ผ๋ฌธ์—์„œ๋Š” ์ดˆ์ ์„ ๋งž์ถ”์—ˆ๋˜ ๋ฌธ์ œ์˜ ํŠน์„ฑ์ƒ ํ•™์Šต์— ๋งŽ์€ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌํ•  ์ˆ˜ ์—†์—ˆ๊ธฐ ๋•Œ๋ฌธ์—ย data augmentation์„ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๋ฐฉ๋ฒ•์œผ๋กœ elastic deformation์„ ์‚ฌ์šฉํ–ˆ๋Š”๋ฐ ๊ฐ pixel์„ gaussian distribution์„ ๊ฐ€์ง€๋Š” random displacement vector๋กœ์˜ ๋ณ€ํ™˜์„ ์ด์šฉํ•ด์„œ ๋ฐ์ดํ„ฐ ์ˆ˜๋ฅผ ์ธ์œ„์ ์œผ๋กœ ๋Š˜๋ ธ๋‹ค๋Š” ์ •๋„๋งŒ ์–ธ๊ธ‰ํ•˜๊ณ  ๋„˜์–ด๊ฐ€๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

Experiments & Results

๋…ผ๋ฌธ์—์„œ ์ œ์‹œํ•œ ์‹คํ—˜๊ณผ ๊ทธ ๊ฒฐ๊ณผ๋Š” ๋…ผ๋ฌธ ์—ฐ๊ตฌ์ง„๋“ค์ด ์ถœ์ „ํ–ˆ๋˜ EM segmentation challenge์™€ ISBI cell tracking challenge์˜ ๊ฒฐ๊ณผ์˜€์Šต๋‹ˆ๋‹ค.
Ranking on the EM segmentation challenge 2015, sorted by Warping Error
Warping error์˜ ๊ด€์ ์—์„œ ๋…ผ๋ฌธ์—์„œ ์ œ์‹œํ•œ U-Net์ด ๊ฐ€์žฅ ์ข‹์€ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” topological ๊ด€์ ์—์„œ์˜ equality๋ฅผ ๋”ฐ์ง„ error์ด๋ฉฐ rand error๋ณด๋‹ค ๊ฒฝ๊ณ„์˜ shifting์— ๋Œ€ํ•ด์„œ ๋ฏผ๊ฐํ•œ error์ž…๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ Pixel Error๋Š” ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ๋Š”, pixel-wise๋กœ ์ œ๋Œ€๋กœ segmentation์ด ์ด๋ฃจ์–ด์กŒ๋Š”์ง€๋ฅผ ํŒ๋‹จํ•˜๋Š” error์ž…๋‹ˆ๋‹ค. ์ž์„ธํ•œ error์— ๋Œ€ํ•œ ์„ค๋ช…์€ย ์ด๊ณณ์„ ์ฐธ๊ณ ํ•˜์‹œ๋ฉด ์ข‹์„ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค.
๋”๋ถˆ์–ด ์ˆ˜์น˜๋กœ ๋˜ ์ œ๊ณตํ•œ ๊ฒƒ์ด IOU(Intersection Over Union)์ž…๋‹ˆ๋‹ค. ์ด๋Š” ground truth์™€ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•จ์— ์žˆ์–ด ๊ต์ฐจ์˜์—ญ์„ ํ•ฉ์˜์—ญ์œผ๋กœ ๋‚˜๋ˆˆ ๊ฐ’์ž…๋‹ˆ๋‹ค. ์ฆ‰, ์–ผ๋งˆ๋‚˜ ์‹ค์ œ์™€ ๋™์ผํ•˜๊ฒŒ ๊ฒน์ณ์ ธ ์žˆ๋Š๋ƒ๋ฅผ ์ธก์ •ํ•˜๋Š” ์ฒ™๋„๋กœ ๋ณด์‹œ๋ฉด ๋ฉ๋‹ˆ๋‹ค.
Segmentation results(IOU) on the ISBI cell tracking challenge 2015
์œ„ ๊ฒฐ๊ณผ์ฒ˜๋Ÿผ ๊ฐ cell๋ณ„๋กœ ๋ณธ ๋…ผ๋ฌธ์ด ๊ตฌํ˜„ํ•œ architecture๊ฐ€ ๊ฐ€์žฅ ๋†’์€ IOU ์ˆ˜์น˜๋ฅผ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์•„๋ž˜ ๊ทธ๋ฆผ์€ ๊ฐ๊ฐ์˜ cell์„ ์‹ค์ œ๋กœ trackingํ•œ ๋ชจ์Šต์ž…๋‹ˆ๋‹ค.
Results on ISBI cell tracking challenge 2015

Conclusion

์ด๊ฒƒ์œผ๋กœ ๋…ผ๋ฌธย โ€œU-Net: Convolutional Networks For Biomedical Image Segmentationโ€์˜ ๋‚ด์šฉ์„ ๊ฐ„๋‹จํ•˜๊ฒŒ ์š”์•ฝํ•ด๋ณด์•˜์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์ด ๋…ผ๋ฌธ์€ ๋…ผ๋ฌธ์˜ ๊ฒฐ๋ก  ๋ถ€๋ถ„์—์„œ U-Net ์ด ๋‹ค์–‘ํ•œ biomedical segmentation application์— ์ ์šฉ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋ผ๊ณ  ์–ธ๊ธ‰ํ•œ ๊ฒƒ์— ๊ธฐ๋Œ€๋ผ๋„ ํ•˜๋Š” ๋“ฏํ•œ ๋†’์€ ์ธ์šฉ์น˜๋ฅผ ์„ ๋ณด์˜€์Šต๋‹ˆ๋‹ค.
๊ฐœ์ธ์ ์œผ๋กœ ๋”ฅ๋Ÿฌ๋‹์ด biomedical ๋ถ„์•ผ์—์„œ ํ˜„์žฌ ์ง„ํ–‰ ์ค‘์ธ ๊ธฐ์—ฌ๋ณด๋‹ค ์•ž์œผ๋กœ ๋” ๋งŽ์€ ๊ฒƒ๋“ค์„ ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ ๊ฐ™์Œ์„ ๋Š๋‚„ ์ˆ˜ ์žˆ์—ˆ๋˜ ๋…ผ๋ฌธ์ด ์•„๋‹ˆ์—ˆ๋‚˜ ์‹ถ์Šต๋‹ˆ๋‹ค.