• ๋Œ€ํ•œ์ „๊ธฐํ•™ํšŒ
Mobile QR Code QR CODE : The Transactions of the Korean Institute of Electrical Engineers
  • COPE
  • kcse
  • ํ•œ๊ตญ๊ณผํ•™๊ธฐ์ˆ ๋‹จ์ฒด์ด์—ฐํ•ฉํšŒ
  • ํ•œ๊ตญํ•™์ˆ ์ง€์ธ์šฉ์ƒ‰์ธ
  • Scopus
  • crossref
  • orcid

  1. (Dept. of AI Convergence Engineering(BK21), Gyeongsang National University, Korea.)
  2. (Dept. of Opthalmology, Gyeongsang National University, Institute of Health Science, Gyeongsang National University Changwon Hospital, Changwon, Korea.)
  3. (School of Computer Science and Engineering, Kyungpook National University, Korea.)



Automated one-hot diagnosis, OCT image, deep learning, Ophthalmic disease classification

1. ์„œ ๋ก 

์ตœ๊ทผ ๊ธฐ๊ณ„ํ•™์Šต ๋ถ„์•ผ์˜ ์—ฐ๊ตฌ๊ฐ€ ๋ฐœ์ „ํ•จ์— ๋”ฐ๋ผ ์ตœ๊ทผ ์˜๋ฃŒ ์˜์ƒ ๋ถ„์•ผ์™€ ๊ธฐ๊ณ„ํ•™์Šต(Machine Learning) ๋ถ„์•ผ์˜ ์œตํ•ฉ์ด ํ™œ๋ฐœํžˆ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋‹ค(1-6). ํŠนํžˆ ์ด๋ฏธ์ง€ ๋ถ„์„ ๋ฐ ๋ถ„๋ฅ˜์— ๊ธฐ๊ณ„ํ•™์Šต ์ ์šฉ์ด ํ™œ๋ฐœํžˆ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ์ฃผ๋กœ ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ ํ™˜์ž์˜ ์—‘์Šค๋ ˆ์ด(X-ray), CT ๊ทธ๋ฆฌ๊ณ  MRI์˜ ์ด๋ฏธ์ง€๋ฅผ ํ™œ์šฉํ•œ ๊ธฐ๊ณ„ํ•™์Šต์ด ๋‚ด๋ฆฐ ์งˆ๋ณ‘ ์ง„๋‹จ, ์˜ˆ์ธก, ์น˜๋ฃŒ ๋ฐฉ๋ฒ• ์ œ์•ˆ ์—ฐ๊ตฌ๊ฐ€ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋‹ค. ๋ฏธ๊ตญ IBM์˜ AI ๋‹ฅํ„ฐ ์™“์Šจ(Watson)์€ ์™ธ๊ตญ์˜ ์•” ์„ผํ„ฐ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ตญ๋‚ด ์•” ์„ผํ„ฐ์—๋„ ๋„์ž…๋˜์–ด ํ์•”, ์ „๋ฆฝ์„ ์•” ๋“ฑ ์•” ์ง„๋ฃŒ๋ฅผ ์ง€์›ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ(7), ์ค‘๊ตญ AI ๋‹ฅํ„ฐ ์ƒค์˜ค์ด(Xiaoyi)๋Š” 2017๋…„์— ์˜์‚ฌ ์ž๊ฒฉ์‹œํ—˜์— ํ•ฉ๊ฒฉํ•ด 2018๋…„ ์ข…ํ•ฉ ๊ฑด๊ฐ•๊ฒ€์ง„ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์žˆ๋‹ค.

์ตœ๊ทผ ์ฝ”๋กœ๋‚˜19์˜ ํŒฌ๋ฐ๋ฏน(pandemic) ์ƒํ™ฉ์—์„œ, ๊ธฐ๊ณ„ํ•™์Šต์— ๊ธฐ๋ฐ˜ํ•œ AI๋ฅผ ํ™œ์šฉํ•œ ๋น„๋Œ€๋ฉด ์ง„๋ฃŒ๊ฐ€ ์ฃผ๋ชฉ์„ ๋ฐ›๊ณ  ์žˆ๋‹ค. ํŠนํžˆ ๋น„๊ต์  ์†์‰ฝ๊ฒŒ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด์„œ ๊ฐ€์ƒ์˜ AI ๋‹ฅํ„ฐ๊ฐ€ 1์ฐจ ํŒ๋‹จ์„ ํ†ตํ•ด์„œ ํ™˜์ž ์งˆ๋ณ‘์˜ ๊ธด๊ธ‰์„ฑ์— ๊ธฐ๋ฐ˜ํ•œ ์šฐ์„ ์ˆœ์œ„๋ฅผ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ œํ•œ๋œ ๋ณ‘์›์˜ ์ธ์ ยท๋ฌผ์  ์ž์›์„ ํ™œ์šฉํ•˜์—ฌ ๊ธด๊ธ‰ํ•œ ํ™˜์ž๋ฅผ ์šฐ์„ ์ ์œผ๋กœ ์น˜๋ฃŒํ•˜๊ฒŒ ํ•จ์œผ๋กœ์จ, ์งˆ๋ณ‘ ์น˜๋ฃŒ์˜ ๊ฐ€๋Šฅ์„ฑ ๋ฟ ์•„๋‹ˆ๋ผ ์ œํ•œ๋œ ์˜๋ฃŒ ์ž์›์„ ํ™œ์šฉ๋„๋ฅผ ๊ทน๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ๋‹ค.

์•ˆ๊ณผ ๋ถ„์•ผ์—์„œ๋Š” ์•ˆ์ €(fundus) ์ด๋ฏธ์ง€์™€ ๊ด‘ ๊ฐ„์„ญ ๋‹จ์ธต ์ดฌ์˜(Optical Coherence Tomography, ์ดํ•˜ OCT) ์ด๋ฏธ์ง€๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์•ˆ๊ณผ ์งˆ๋ณ‘์„ ์ง„๋‹จํ•˜๊ณ  ์น˜๋ฃŒ ์˜ˆํ›„๋ฅผ ์˜ˆ์ธกํ•˜๊ณ  ์žˆ๋‹ค. ๊ตฌ๊ธ€ ์‚ฐํ•˜ ๋”ฅ๋งˆ์ธ๋“œ(DeepMind)๊ฐ€ ๊ฐœ๋ฐœํ•œ AI ์‹œ์Šคํ…œ์€ ๋ง๋ง‰ ์Šค์บ” ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•ด ์•ˆ๊ตฌ์˜ ์งˆํ™˜ ๋ฐ ์ง•ํ›„๋ฅผ 94.5%์˜ ๋†’์€ ์ •ํ™•๋„๋กœ ์ง„๋‹จํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํ–ฅํ›„ ์‹ค์ œ ํ™˜์ž ์ง„๋‹จ์—๋„ ์ ์šฉ๋˜๊ณ  ์žˆ๋‹ค.

OCT๋Š” ๊ทผ์ ์™ธ์„ ์„ ์‚ฌ์šฉํ•˜์—ฌ, ๊ด‘ํ•™ ์‚ฐ๋ž€ ๋งค์ฒด๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋งˆ์ดํฌ๋กœ๋ฏธํ„ฐ ํ•ด์ƒ๋„์˜ 3์ฐจ์› ์ด๋ฏธ์ง€๋ฅผ ์บก์ฒ˜ํ•˜๋Š” ์˜๋ฃŒ ์˜์ƒ ๊ธฐ์ˆ ์ด๋‹ค. ์•ˆ๊ณผ, ํ”ผ๋ถ€๊ณผ, ์‹ฌ์žฅํ•™ ๋“ฑ์˜ ์—ฌ๋Ÿฌ ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ํŠนํžˆ ๋งˆ์ดํฌ๋กœ๋ฏธํ„ฐ ํ•ด์ƒ๋„์™€ ๋ฐ€๋ฆฌ๋ฏธํ„ฐ ์นจํˆฌ ๊นŠ์ด๊ฐ€ ํ•„์š”ํ•œ ์•ˆ๊ณผ์šฉ ๋ฐ ๊ธฐํƒ€ ์กฐ์ง ์ด๋ฏธ์ง•์— ์ ํ•ฉํ•˜๋‹ค. ์•ˆ๊ณผ์—์„œ๋Š” ์ง„๋ฃŒํ•˜๊ธฐ ์ „ ๋ง๋ง‰ ๋‚ด์—์„œ ์ƒ์„ธํ•œ ์ด๋ฏธ์ง€๋ฅผ ์–ป์Œ์œผ๋กœ, ๋ง๋ง‰ ๊ตฌ์กฐ ๋ฐ ์ƒํƒœ๋ฅผ ์‹œ๊ฐํ™”ํ•˜์—ฌ ๋…ธํ™”์„ฑ ํ™ฉ๋ฐ˜ ๋ณ€์„ฑ(Age-related Macular Degeneration, ์ดํ•˜ AMD), ๋‹น๋‡จ์„ฑ ํ™ฉ๋ฐ˜๋ถ€์ข…(Diabetic Macular Edema, ์ดํ•˜ DME), ๋“œ๋ฃจ์  (Drusen), ๋‹น๋‡จ ๋ง๋ง‰์ฆ(Diabetic Retinopathy) ๋˜๋Š” ๋งฅ๋ฝ๋ง‰ ์‹ ์ƒํ˜ˆ๊ด€(Choroidal Neovascularization, ์ดํ•˜ CNV)๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์˜ํ•™์  ์ง„๋‹จ์„ ๋‚ด๋ฆฌ๋Š”๋ฐ ์ค‘์š”ํ•œ ์ •๋ณด๋กœ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ๋‹ค.

๊ทธ๋ฆผ. 1. OCT ์ด๋ฏธ์ง€: (a) AMD (b) DME (c) Drusen

Fig. 1. OCT Image: (a) AMD (b) DME (c) Drusen

../../Resources/kiee/KIEE.2021.70.7.1036/fig1.png

๊ทธ๋ฆผ 1์—์„œ๋Š” AMD, DME ๊ทธ๋ฆฌ๊ณ  Drusen์˜ ๋ณ‘์ฆ์˜ OCT ์ด๋ฏธ์ง€๋ฅผ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ์งˆ๋ณ‘์— ๋Œ€ํ•ด์„œ๋„ ๊ธฐ๊ณ„ํ•™์Šต ๊ธฐ๋ฐ˜์˜ ์ธ๊ณต์ง€๋Šฅ์ด ํ™œ๋ฐœํžˆ ์ ์šฉ๋˜๊ณ  ์žˆ๋Š”๋ฐ, ํŠนํžˆ ์‹ฌ์ธต CNN(Convolution Neural Network)๊ณผ ์ „์ดํ•™์Šต(Transfer Learning)์„ ํ™œ์šฉํ•˜์—ฌ ์ ์€ ๋ฐ์ดํ„ฐ ์ˆ˜๋กœ๋„ ํ•™์Šต์„ ๋น ๋ฅด๊ฒŒ ํ•˜๋ฉฐ, ๋†’์€ ์˜ˆ์ธก ์ •ํ™•๋„๋ฅผ ๊ฑฐ๋‘๊ณ  ์žˆ๋‹ค(1). ํŠนํžˆ DenseNet121(8)์„ ํ†ตํ•ด์„œ๋Š” OCT ์ด๋ฏธ์ง€ ์ƒ์—์„œ ์œ„์˜ 3๊ฐ€์ง€ ์งˆ๋ณ‘ AMD, DME ๊ทธ๋ฆฌ๊ณ  Drusen์— ๋Œ€ํ•ด์„œ 95% ์ด์ƒ์˜ ์ •ํ™•๋„๋ฅผ ๋ณด์ด๊ณ  ์žˆ๋‹ค.

ํ•˜์ง€๋งŒ ์‹ค์ œ ํ™˜์ž์˜ ์•ˆ๊ตฌ์—์„œ OCT ์ด๋ฏธ์ง€๋ฅผ ์ดฌ์˜ํ•  ๋•Œ์—๋Š” ๋‹จ์ผ ์ด๋ฏธ์ง€๊ฐ€ ์•„๋‹ˆ๋ผ, ๊ทธ๋ฆผ 2์™€ ๊ฐ™์ด ๋‹ค์ˆ˜(20์—ฌ ์žฅ)์˜ ์‚ฌ์ง„์ด ์ดฌ์˜๋˜๋Š”๋ฐ ์ด๋Ÿฌํ•œ ์‚ฌ์ง„๋“ค์— ๋Œ€ํ•œ ํŒ๋…์„ ๊ธฐ์กด์˜ ๊ฐ€์žฅ ์ •ํ™•๋„๊ฐ€ ๋†’์€ DenseNet121๋กœ ์ˆ˜ํ–‰ํ•˜์—ฌ๋„ ๋ถˆ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ƒ์ถฉ๋˜๋Š” ๊ฒฐ๊ณผ๋“ค์ด ๋Œ€๋ถ€๋ถ„์ด๋‹ค. ์ „๋ฌธ์˜์˜ ํŒ๋… ์—ญ์‹œ ๊ฐ OCT ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ํŒ๋…์€ ์„œ๋กœ ๋ถˆ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ƒ์ถฉ๋  ์ˆ˜ ์žˆ์œผ๋‚˜, ์ „๋ฌธ์˜์˜ ํ•™์Šต๊ณผ ๊ฒฝํ—˜์„ ํ†ตํ•ด ๋šœ๋ ทํ•œ ํ•œ๋‘ ๊ฐ€์ง€ ๊ฒฐ๊ณผ๋ฅผ ์ง€๋ชฉํ•˜์—ฌ ์น˜๋ฃŒ ๋ฐฉ๋ฒ•์„ ์„ ํƒํ•œ๋‹ค. ํ•˜์ง€๋งŒ ์—ฌ์ „ํžˆ ๊ธฐ๊ณ„ํ•™์Šต ๊ธฐ๋ฐ˜์˜ ์ธ๊ณต์ง€๋Šฅ์€ ์ผ๋ฐ˜ ์•ˆ๊ณผ ์ „๋ฌธ์˜๋“ค์ด ์ˆ˜ํ–‰ํ•˜๋Š” ๋‹ค์ค‘ OCT ์ด๋ฏธ์ง€๋ฅผ ์‚ฌ์šฉํ•œ ๋‹จ์ผ ๊ฒฐ๊ณผ๋ฅผ ๋„์ถœํ•˜์ง€ ๋ชปํ•˜๊ณ  ์žˆ๋‹ค.

๊ทธ๋ฆผ. 2. OCT ์ด๋ฏธ์ง€ ์ดฌ์˜

Fig. 2. OCT Image Acquisition

../../Resources/kiee/KIEE.2021.70.7.1036/fig2.png

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹จ์ผ OCT ์ด๋ฏธ์ง€๋“ค์„ ๊ฐ€์ง€๊ณ  ์œ„์˜ 3๊ฐœ ์งˆํ™˜์— ๋Œ€ํ•ด ์ „๋ฌธ์˜์˜ ์ง„๋‹จ๊ณผ ์œ ์‚ฌํ•œ ์ข…ํ•ฉ์ ์ธ ํŒ๋‹จ์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋Š” ๋‹ค์–‘ํ•œ ๊ธฐ๊ณ„ํ•™์Šต ํ”„๋ ˆ์ž„์›Œํฌ ๋ชจ๋ธ์„ ์ œ์‹œํ•˜๊ณ  ๊ทธ์— ๋Œ€ํ•œ ์ ์šฉ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•œ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์œ„ ์งˆํ™˜์— ๋Œ€ํ•œ 2๋‹จ๊ณ„์˜ ์‹ฌ์ธต ๊ธฐ๊ณ„ํ•™์Šต ๋ฐฉ๋ฒ•์„ ๋ณด์ด๊ณ ์ž ํ•œ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ๊ธฐ์กด์˜ ์‹ฌ์ธต CNN ๊ธฐ๋ฐ˜์˜ DenseNet121์„ ์‚ฌ์šฉํ•œ ๋‹จ์ผ OCT ์ด๋ฏธ์ง€ ์งˆ๋ณ‘ ๋ถ„๋ฅ˜ ์ˆ˜ํ–‰ํ•œ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ์‹ค์ œ ์•ˆ๊ณผ ์ „๋ฌธ์˜๋“ค์ด ์ง„๋‹จ ๋‚ด๋ฆฌ๋Š” ๊ฒฐ๊ณผ๋ฅผ ํ•™์Šตํ•˜์—ฌ, ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์˜ ๋ถ„๋ฅ˜ํ•œ ๋‹ค์ˆ˜์˜ OCT ์ด๋ฏธ์ง€๋“ค์— ๋Œ€ํ•œ ์ข…ํ•ฉ์ ์ธ ํŒ๋‹จ์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋Š” ์‹œ์Šคํ…œ์„ ์ œ์•ˆํ•œ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ง‘๋‹จ ๋ถ„๋ฅ˜๋ฅผ ์œ„ํ•ด 3๊ฐ€์ง€ ์‹ฌ์ธต ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์ธ FC(Fully Connected), LSTM(Long Short-Term Memory) ๊ทธ๋ฆฌ๊ณ  CNN์œผ๋กœ ๊ตฌํ˜„๋œ ๋ชจ๋ธ์„ ์†Œ๊ฐœํ•˜๊ณ  ๊ฐ๊ฐ์˜ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๋ถ„๋ฅ˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€๋‹ค.

๋ณธ ๋…ผ๋ฌธ์˜ ๊ตฌ์„ฑ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค. 2์žฅ์—์„œ๋Š” OCT ์ด๋ฏธ์ง€๋ฅผ ํ™œ์šฉํ•œ ๊ด€๋ จ ์—ฐ๊ตฌ ๊ธฐ์ˆ  ๋ฐ ์‹ฌ์ธตํ•™์Šต์˜ ์ž๋™ ์ง„๋‹จ๊ณผ ๊ด€๋ จ๋œ ๊ธฐ์กด ์—ฐ๊ตฌ์— ๋Œ€ํ•˜์—ฌ ์†Œ๊ฐœํ•œ๋‹ค. 3์žฅ์—์„œ๋Š” ๋ณธ ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•œ ๋ฐฉ๋ฒ•์ธ DenseNet121 ๋ฐ ๋‹ค์–‘ํ•œ ๊ธฐ๊ณ„ํ•™์Šต ๋ชจ๋ธ์„ ํ†ตํ•ด ์ž๋™ ์ง„๋‹จ ์‹œ์Šคํ…œ์„ ์†Œ๊ฐœํ•œ๋‹ค. 4์žฅ์—์„œ๋Š” ์ž๋™ ์ง„๋‹จ ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ถ„์„ํ•˜๊ณ , ๋งˆ์ง€๋ง‰์œผ๋กœ 5์žฅ์—์„œ๋Š” ๊ฒฐ๋ก ์— ๋Œ€ํ•ด ๊ธฐ์ˆ ํ•œ๋‹ค.

2. ๊ด€๋ จ ์—ฐ๊ตฌ

2.1 OCT ์ด๋ฏธ์ง€๋ฅผ ํ™œ์šฉํ•œ ๊ด€๋ จ ์—ฐ๊ตฌ ๊ธฐ์ˆ 

์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ๊ธฐ๊ณ„ํ•™์Šต์ด ๋ฐœ์ „ํ•˜๋ฉด์„œ ์˜๋ฃŒ ์˜์ƒ์„ ํ™œ์šฉํ•œ ๋‹ค์–‘ํ•œ ์‹ฌ์ธตํ•™์Šต ์—ฐ๊ตฌ๊ฐ€ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋‹ค(1-6). ์•ˆ๊ณผ ์˜๋ฃŒ ์˜์ƒ์—๋Š” ๋Œ€ํ‘œ์ ์œผ๋กœ ์•ˆ์ € ์ด๋ฏธ์ง€์™€ OCT ์ด๋ฏธ์ง€๊ฐ€ ์žˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ๋“ค์„ ํ™œ์šฉํ•œ ์—ฐ๊ตฌ์—๋Š” ๋Œ€ํ‘œ์ ์œผ๋กœ ์˜๋ฃŒ ์˜์ƒ ๋‚ด์—์„œ ์‹œ๊ฐ์ ์œผ๋กœ ๋“œ๋Ÿฌ๋‚˜๋Š” ์งˆ๋ณ‘ ํŠน์ง•์„ ํ†ตํ•œ ์งˆ๋ณ‘ ์ง„๋‹จ์ด ์žˆ๋‹ค. OCT ๊ด€๋ จ ๋Œ€ํ‘œ ์—ฐ๊ตฌ๋กœ์„œ, OCT ์ด๋ฏธ์ง€ ๋‚ด์˜ ์ˆ˜์•ก๊ณผ ๋ง๋ง‰ ์ธต ๋ถ„ํ• (segmentation) ์—ฐ๊ตฌ(2), OCT ์ด๋ฏธ์ง€์—์„œ ์งˆ๋ณ‘ ์ฆ์ƒ์ด ๋ณด์ด๋Š” ํŠน์ง• ๋ถ€๋ถ„ ๊ฐ์ง€(detection) ์—ฐ๊ตฌ(3,6)๊ฐ€ ์žˆ๋‹ค. ๋˜ํ•œ, ๊ฐ๋ง‰๊ณผ ๋ง๋ง‰ ์‚ฌ์ด์˜ ๋งฅ๋ฝ๋ง‰ ์˜์—ญ์˜ ๋ฌด๋Šฌ, 2์ฐจ์› ํ˜น์€ 3์ฐจ์›์˜ ํ˜•ํƒœ์ (morphological) ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ณ  ์ด๋ฅผ ์•ˆ๊ณผ์  ์งˆ๋ณ‘ ์ง„๋‹จ์— ํ™œ์šฉํ•˜๋Š” ๋‹ค์–‘ํ•œ ์—ฐ๊ตฌ๋“ค์ด ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋‹ค(9).

2.2 ์‹ฌ์ธตํ•™์Šต์˜ ์ž๋™ ์ง„๋‹จ ์—ฐ๊ตฌ

2018๋…„ cell์€ ๋”ฅ ๋Ÿฌ๋‹์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์•ˆ๊ตฌ ์งˆํ™˜ ์ง„๋‹จ ์‹œ์Šคํ…œ์— ๋Œ€ํ•œ Kermany et a.l.(1)์˜ ๋…ผ๋ฌธ์„ ๋ฐœํ‘œํ–ˆ์œผ๋ฉฐ, Kermany et al.(1)์—์„œ๋Š” Inception V3 ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ OCT ์ด๋ฏธ์ง€ ํ•œ ์žฅ์— ๋Œ€ํ•ด์„œ CNV, DME, Drusen๊ณผ ์ •์ƒ์„ ๋ถ„๋ฅ˜ํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€์œผ๋ฉฐ, ๊ทธ ์„ฑ๋Šฅ์€ 99%์˜ ์ •ํ™•๋„๋กœ ํƒ์›”ํ–ˆ๋‹ค.

Kaymak et al.(6)๋Š” OCT ์ด๋ฏธ์ง€์—์„œ AMD ๋ฐ DME์˜ ์ž๋™ ๊ฐ์ง€์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋ฅผ ๋ฐœํ‘œํ–ˆ์œผ๋ฉฐ, AMD์ค‘์—์„œ ๊ฑด์„ฑ AMD, ์Šต์‹ AMD๋ฅผ ๋ถ„๋ฅ˜ํ•˜๊ณ  DME๋ฅผ ํฌํ•จํ•˜์—ฌ ์ด 3๊ฐ€์ง€์˜ ์งˆ๋ณ‘์„ ๋ถ„๋ฅ˜ํ•˜์˜€๋‹ค. ์ด ์—ฐ๊ตฌ ๋˜ํ•œ OCT ์ด๋ฏธ์ง€๋ฅผ AMD ๋ฐ DME ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ ์ œ์•ˆ๋œ AlexNet์˜ ์ „์ด ํ•™์Šต ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์˜€์œผ๋ฉฐ ๊ทธ ๊ฒฐ๊ณผ 97%์ด์ƒ์˜ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•˜์˜€๋‹ค.

Cecilia et al.(5)๋Š” AMD ํ™˜์ž์˜ OCT ์ด๋ฏธ์ง€์™€ ์ •์ƒ OCT ์ด๋ฏธ์ง€๋ฅผ ๊ตฌ๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด ์‹ฌ์ธต ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์ธ VGG16 ๋ชจ๋ธ์— Xavier ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ€์ค‘์น˜๋ฅผ ์ดˆ๊ธฐํ™”ํ•˜์—ฌ ๋ชจ๋ธ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” 93.45%์˜ ์ •ํ™•๋„๋กœ 97.45%์˜ AUC๋ฅผ ๋‹ฌ์„ฑํ•˜์˜€๋‹ค.

Xiaogang et al.(4)๋Š” ์‚ฌ์ „ ํ›ˆ๋ จ ๋œ ImageNet๊ณผ ์ง€์› ๋ฒกํ„ฐ ๋จธ์‹ (support vector machine)์„ ์‚ฌ์šฉํ•˜์—ฌ ์•ˆ์ € ์ด๋ฏธ์ง€๋กœ ๋ถ€ํ„ฐ ๋‹น๋‡จ๋ณ‘ ์„ฑ ๋ง๋ง‰์ฆ(DR, Diabetic Retinopathy)์„ ๋ถ„๋ฅ˜ํ•˜๋Š” ๋…ผ๋ฌธ์„ ๋ฐœํ‘œํ–ˆ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์‚ฌ์ „ ํ›ˆ๋ จ๋œ CNN ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์•ˆ์ € ์ด๋ฏธ์ง€์—์„œ ํŠน์ง•์„ ์ถ”์ถœํ•˜์˜€๊ณ  ์ด๋Ÿฌํ•œ ํŠน์ง• ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฐ’์„ ํ†ตํ•ด ์ง€์› ๋ฒกํ„ฐ ๋จธ์‹ (support vector machine)์„ ํ›ˆ๋ จํ•˜์—ฌ ์ตœ์ข… ์ง„๋‹จ ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜์˜€๋‹ค.

๋˜ํ•œ, Wang et al.(3)๋Š” 2๊ฐœ์˜ CNN๋ชจ๋ธ์„ ํ†ตํ•ด ๋ฏธ์ˆ™์•„ ๋ง๋ง‰๋ณ‘์ฆ(ROP, Retinopathy of Prematurity)์„ ์ž๋™ ์ง„๋‹จํ•˜๋Š” ๋…ผ๋ฌธ์„ ๋ฐœํ‘œํ–ˆ๋‹ค. 5511๊ฐœ์˜ ๋ง๋ง‰ ์‚ฌ์ง„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ 5 ๋ถ„ํ•  ๊ต์ฐจ ๊ฒ€์ •์„ ์‚ฌ์šฉํ•˜์—ฌ CNN์„ ํ›ˆ๋ จํ•˜์˜€์œผ๋ฉฐ, 100๊ฐœ์˜ ์ด๋ฏธ์ง€๋กœ ๊ตฌ์„ฑ๋œ ๋…๋ฆฝ์ ์ธ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ 91%์˜ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ ๋ฏธ์ˆ™์•„ ์ „๋ฌธ๊ฐ€์˜ ๋ง๋ง‰ ๋ณ‘์ฆ 8๋ช… ์ค‘ 6๋ช…์„ ๋Šฅ๊ฐ€ํ–ˆ๋‹ค. ๊ทธ ์™ธ์— ๋งŽ์€ ์งˆ๋ณ‘ ํŒ๋‹จ ๋ชจ๋ธ AI ๋ชจ๋ธ์ด ๊ฐœ๋ฐœ๋˜๊ณ  ์žˆ๋‹ค.

๋ณธ ๋…ผ๋ฌธ์€ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์„ ํ–‰ ์—ฐ๊ตฌ๋“ค์„ ์ฐธ๊ณ ํ•˜์—ฌ, Kermany et al.(1)์˜ ๋‹จ์ผ OCT ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ์ง„๋‹จ์€ ํ™˜์ž์—๊ฒŒ ์ง„๋‹จ์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ๋Š” ๋ฌธ์ œ๋กœ ํŒ๋‹จํ•˜์˜€๋‹ค. ์ด์— ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋‹ค์ˆ˜์˜ OCT ์ด๋ฏธ์ง€๋“ค์— ๋Œ€ํ•œ ์ข…ํ•ฉ์ ์ธ ํŒ๋‹จ์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋Š” 2๋‹จ๊ณ„์˜ ์‹ฌ์ธต ๊ธฐ๊ณ„ํ•™์Šต ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•œ๋‹ค. 1๋‹จ๊ณ„์—์„œ DenseNet121์„ ํ™œ์šฉํ•œ OCT ์ด๋ฏธ์ง€ ๋ถ„์„ ์—ฐ๊ตฌ๋ฅผ ์ œ์•ˆํ•œ๋‹ค. 2๋‹จ๊ณ„์—์„œ๋Š” ์—ฌ๋Ÿฌ OCT ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ ์ข…ํ•ฉ์ ์ธ ์ง„๋‹จ ํ›„, ํ™˜์ž์—๊ฒŒ ์•ˆ๊ตฌ ์งˆํ™˜ ์ง„๋‹จํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ œ์•ˆํ•œ๋‹ค. ๋˜ํ•œ, ์„ฑ๋Šฅ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๊ธฐ๊ณ„ํ•™์Šต ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ ๋น„๊ต ๊ฒฐ๊ณผ๋„ ์ œ์‹œํ•œ๋‹ค.

3. ์‹ฌ์ธตํ•™์Šต ๊ธฐ๋ฒ•์„ ์ด์šฉํ•œ ์•ˆ๊ตฌ ์งˆํ™˜ ์ง„๋‹จ

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์•ˆ๊ตฌ ์งˆํ™˜ ์ž๋™ํ™” ์ง„๋‹จ ์‹œ์Šคํ…œ์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด, ์ตœ๊ทผ ์ด๋ฏธ์ง€ ์˜์ƒ ์ฒ˜๋ฆฌ์— ๋Œ€ํ•˜์—ฌ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๊ณ  ์žˆ๋Š” ์‹ฌ์ธตํ•™์Šต ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜์˜€๋‹ค. ํŠนํžˆ OCT ์ด๋ฏธ์ง€์—์„œ ์•ˆ๊ตฌ ์งˆํ™˜์˜ ์‹œ๊ฐ์  ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด DenseNet์„ ์ ์šฉํ•˜์˜€์œผ๋ฉฐ, ์ถ”์ถœ๋œ ํŠน์ง•์„ ํ†ตํ•ด ๋‹ค์–‘ํ•œ ์‹ฌ์ธตํ•™์Šต ๊ธฐ๋ฒ•๋“ค์„ ์ ์šฉ์‹œ์ผœ ํ™˜์ž์—๊ฒŒ ์ตœ์ข…์ ์œผ๋กœ ์˜ํ•™์  ๊ฒฐ์ •์„ ๋ณด์กฐํ•  ์ˆ˜ ์žˆ๋Š” ์ž๋™ํ™” ์ง„๋‹จ ๋ชจ๋ธ์„ 2๊ฐ€์ง€ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค. ๋ณธ ์žฅ์—์„œ๋Š” ์ œ์•ˆํ•œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ตฌํ˜„ ์„ค๋ช…์„ ์œ„ํ•˜์—ฌ ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ์™€ ๊ฐ ๋‹จ๊ณ„์—์„œ ์ ์šฉ๋œ ์‹ฌ์ธตํ•™์Šต ๋ชจ๋ธ์˜ ๊ตฌ์กฐ๋ฅผ ๊ฐ ์ ˆ๋กœ ๋‚˜๋ˆ„์–ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์†Œ๊ฐœํ•œ๋‹ค.

3.1 ๋ฐ์ดํ„ฐ ์…‹ ๊ตฌ์„ฑ

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” OCT ์ด๋ฏธ์ง€๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์•ˆ๊ตฌ ์งˆํ™˜ ์ž๋™ํ™” ์ง„๋‹จ ์‹œ์Šคํ…œ์„ ๊ตฌํ˜„ํ•˜์˜€๋‹ค. ์ž๋™ํ™” ์ง„๋‹จ ์‹œ์Šคํ…œ์˜ ๊ตฌ์กฐ๋Š” ๊ทธ๋ฆผ 3๊ณผ ๊ฐ™์ด ๋‘ ๊ฐ€์ง€์˜ ๋‹จ๊ณ„๋กœ ๋ชจ๋ธ์ด ๊ตฌ์„ฑ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๊ฐ ๋‹จ๊ณ„์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ ์…‹ ๊ตฌ์„ฑ์— ์†Œ๊ฐœํ•œ๋‹ค.

์ „๋ฌธ ๊ฒ€์‚ฌ๊ด€์ด Spectralis OCT ์žฅ์น˜(๋…์ผ ํ•˜์ด๋ธ๋ฒ ๋ฅด๊ทธ ์†Œ์žฌ์˜ ํ•˜์ด๋ธ๋ฒ ๋ฅด๊ทธ ์—”์ง€๋‹ˆ์–ด๋ง)๋กœ ๋ง๋ง‰ OCT๋ฅผ ์ดฌ์˜ํ•˜์˜€์œผ๋ฉฐ, ์ด ์‹œ์Šคํ…œ์€ ์กฐ์ง์—์„œ 3.9m ํ”ฝ์…€์˜ ์ถ• ํ•ด์ƒ๋„์™€ 5.7m ํ”ฝ์…€์˜ ๊ฐ€๋กœ ํ•ด์ƒ๋„๋กœ ์ดˆ๋‹น 40000๊ฐœ์˜ ์Šค์บ”์„ ์‹œํ–‰ํ•˜์˜€๋‹ค. ํ™ฉ๋ฐ˜์˜ ์ˆ˜ํ‰ ๋‹จ๋ฉด ์˜์ƒ ๋˜๋Š” ์ˆ˜์ง ๋‹จ๋ฉด ์˜์ƒ์„ ์ดฌ์˜ํ–ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” AutoIt๋กœ ์ž‘์„ฑ๋œ ์ž๋™ํ™” ํ”„๋กœ๊ทธ๋žจ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€๋ฅผ BMP ํŒŒ์ผ ํ˜•์‹์œผ๋กœ ์ด๋ฏธ์ง€๋ฅผ ์ˆ˜์ง‘ํ•˜์˜€๋‹ค. OCT ์ด๋ฏธ์ง€๋Š” ๊ทธ๋ฆผ 2์™€ ๊ฐ™์ด ์ดฌ์˜ํ•˜์—ฌ AMD, DME, Drusen ํ™˜์ž๋กœ๋ถ€ํ„ฐ ์ˆ˜์ง‘ํ•˜์˜€๋‹ค. ๋˜ํ•œ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ๋ฐ์ดํ„ฐ๋Š” ์‚ญ์ œํ›„ ๋ฐ์ดํ„ฐ ์ฆ๋Œ€๋ฅผ ํ†ตํ•ด์„œ ๋งค ๋ฐ์ดํ„ฐ ์…‹์„ 27์žฅ์œผ๋กœ ์ผ์น˜์‹œ์ผœ์„œ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” 2๊ฐ€์ง€์˜ ๋‹จ๊ณ„๋กœ ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•˜์˜€์œผ๋ฉฐ, ์ฒซ ๋‹จ๊ณ„๋Š” ๊ฐœ๋ณ„ ๋ถ„๋ฅ˜๋กœ ๋‹จ์ผ OCT ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ ์งˆ๋ณ‘์˜ ํŠน์ง•์„ ์ถ”์ถœํ•˜์—ฌ ๊ฐ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ์•ˆ๊ตฌ ์งˆํ™˜์„ ์ง„๋‹จ ๋‚ด๋ฆฌ๋Š” ๋ชจ๋ธ์„ ๊ตฌํ˜„ํ•˜์˜€๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์˜ ๋ชจ๋ธ์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ๋Š” cell์—์„œ ๋ฐœํ‘œ๋œ Kermany et al.(1)์˜ ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ ์…‹์„ ์‚ฌ์šฉํ•˜์˜€๊ณ , ๋ฐ์ดํ„ฐ ์…‹์€ ๋ผ๋ฒจ์ด ์™„๋ฃŒ๋œ ๊ฐ๊ฐ์˜ ์ด๋ฏธ์ง€๊ฐ€ ์ด 84,484์žฅ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๋‹ค. ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ ์…‹์˜ 83,484์žฅ์œผ๋กœ ํ›ˆ๋ จ์„ ์‹œํ‚ค๊ณ , ๋‚˜๋จธ์ง€ 1,000์žฅ์œผ๋กœ ๋ณธ ๋ชจ๋ธ์˜ ํ…Œ์ŠคํŠธ๋ฅผ ์ง„ํ–‰ํ•˜์˜€๋‹ค. ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ํ†ตํ•ด ์•ˆ๊ณผํ•™์—์„œ ์ œ๊ณต๋œ OCT ์ด๋ฏธ์ง€์— ๋Œ€ํ•ด์„œ ๋ผ๋ฒจ ์ž‘์—…์ด ์ด๋ฃจ์–ด์กŒ๋‹ค. OCT ์ด๋ฏธ์ง€๋Š” ๊ฐ๊ฐ AMD 27,675์žฅ, DME 25,137์žฅ๊ณผ Drusen 18,684์žฅ์œผ๋กœ ์ด 71,496์žฅ์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๋‹ค.

๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š” ์ง‘๋‹จ ๋ถ„๋ฅ˜๋กœ์จ, ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์˜ ๋ชจ๋ธ์„ ํ†ตํ•ด ๊ฐ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ๋ผ๋ฒจ์ด ์™„๋ฃŒ๋œ 71,496์žฅ์„ ํ†ตํ•ด 27๊ฐœ๋ฅผ ํ•œ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์…‹์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ์‹ ๊ฒฝ๋ง์ด ์ข…ํ•ฉ์ ์œผ๋กœ ํ•˜๋‚˜์˜ ์ง„๋‹จ์„ ํ™˜์ž์—๊ฒŒ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋„๋ก ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ๋ฐ์ดํ„ฐ ์…‹์˜ ํ–‰์€ ํ•œ ์žฅ์˜ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ AMD, DME, Drusen, ์ •์ƒ์˜ ๊ฐ ํ™•๋ฅ ์„ ์˜๋ฏธํ•˜๋ฉฐ, ์—ด์€ 27์žฅ์˜ ์ด๋ฏธ์ง€๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” 27๊ฐœ์˜ ๋ณผ๋ฅจ ๋ฐ์ดํ„ฐ ํ˜•์‹์œผ๋กœ ๊ฐ๊ฐ AMD 1,025๊ฐœ, DME 931๊ฐœ์™€ Drusen 692๊ฐœ๋กœ ์ด 2,648๊ฐœ๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋‹ค.

3.2 ๋‹จ์ผ OCT ์ด๋ฏธ์ง€ ์ง„๋‹จ ๊ฒฐ์ • ๋ชจ๋ธ

๊ทธ๋ฆผ. 3. OCT ์ด๋ฏธ์ง€๋ฅผ ์œ„ํ•œ 2-Phase ์ž๋™ํ™” ์ง„๋‹จ ์‹œ์Šคํ…œ

Fig. 3. 2-Phase Automated Diagnostic System for OCT-images

../../Resources/kiee/KIEE.2021.70.7.1036/fig3.png

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒˆ๋กœ์šด ์‹ ๊ฒฝ๋ง ๊ตฌ์กฐ๋ฅผ ์„ค๊ณ„ํ•˜๋Š” ๋Œ€์‹  ๊ฒฐ์ • ๋ชจ๋ธ์˜ ์ „๋ฐฉ์—๋Š” DensNet121 ๋„คํŠธ์›Œํฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์„ค๊ณ„ํ•˜์˜€๊ณ , ๊ฒฐ์ • ๋ชจ๋ธ์˜ ํ›„๋ฐฉ์—๋Š” ๋‘ ๊ฐœ์˜ FC ์ธต๋“ค๋กœ ์„ค๊ณ„ํ•˜์˜€๋‹ค. ํ›„๋ฐฉ์˜ ์ฒซ ๋ฒˆ์งธ FC ์ธต์€ DenseNet121์œผ๋กœ๋ถ€ํ„ฐ ์ „๋‹ฌ๋ฐ›์€ ์ •๋ณด์˜ ์ •๋ณด ์†์‹ค์„ ์ตœ์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์œผ๋ฉฐ, ์ง€์—ญ์  ์ตœ์ (local optimum) ๋ฌธ์ œ๋ฅผ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ๋ฐฐ์น˜ ์ •๊ทœํ™”(Batch-Normalization)์™€ ReLU(rectification) ํ™œ์„ฑํ™” ํ•จ์ˆ˜๊ฐ€ ์‚ฌ์šฉ๋œ๋‹ค. ์ตœ์ข… FC ์ธต์€ ๊ฐ์ฒด(Object)๋ฅผ ๋ถ„๋ฅ˜ํ•˜๊ธฐ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์œผ๋ฉฐ, softmax ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข…์ ์œผ๋กœ AMD, DME, Drusen ๋ฐ ์ •์ƒ ์ผ ํ™•๋ฅ ์˜ ํ˜•ํƒœ๋กœ ๊ฒฐ๊ณผ(output)๋ฅผ ์–ป๊ฒŒ ๋œ๋‹ค.

์ž…๋ ฅ ์ธต(input)์—์„œ๋Š” ๋‹จ์ผ OCT ์ด๋ฏธ์ง€๊ฐ€ ํ‘๋ฐฑ ์ฒ˜๋ฆฌ๋˜์–ด 3์ฐจ์›์˜ ํ˜•ํƒœ๋กœ ์ž…๋ ฅ๋˜๋ฉฐ, ์ž…๋ ฅ ํฌ๊ธฐ(input size)๋Š” (None, 299, 299, 3)์ด๋‹ค.

๊ทธ๋ฆฌ๊ณ  ์ถœ๋ ฅ ์ธต(output)์€ softmax ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์ข…์ ์œผ๋กœ AMD, DME, Drusen ๋ฐ ์ •์ƒ ์ผ ํ™•๋ฅ ์˜ ํ˜•ํƒœ๋กœ ๊ฒฐ๊ณผ(output)๋ฅผ ์–ป๊ธฐ ๋•Œ๋ฌธ์— ์ถœ๋ ฅ ํฌ๊ธฐ(output size)๋Š” (None, 4)์ด๋‹ค.

DenseNet121์€ ์ด 4๊ฐœ์˜ dense block์œผ๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋‹ค. dense block์€ ๋ชจ๋“  ๋ ˆ์ด์–ด๋ฅผ ์„œ๋กœ ์ง์ ‘์ ์œผ๋กœ ์—ฐ๊ฒฐํ•˜์—ฌ ์ปจ๋ฒŒ๋ฃจ์…˜ ์‹ ๊ฒฝ๋ง์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๋ชจ๋“ˆ๋กœ์จ, 1x1 ๋ฐ 3x3 ํฌ๊ธฐ์˜ ์ปค๋„์ด ์žˆ๋Š” ๋‘ ๊ฐœ๊ฐ€ ์žˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ฒซ ๋ฒˆ์งธ dense block์€ 6ํšŒ, ๋‘ ๋ฒˆ์งธ๋Š” 12ํšŒ, ์„ธ ๋ฒˆ์งธ๋Š” 24ํšŒ, ๋งˆ์ง€๋ง‰ dense block์—์„œ๋Š” 16ํšŒ๋ฅผ ๋ฐ˜๋ณตํ•œ๋‹ค.

ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋œ ์†์‹ค ํ•จ์ˆ˜(loss function)๋Š” ๋‹ค์ค‘ ๋ถ„๋ฅ˜ ์†์‹ค ํ•จ์ˆ˜(categorical cross-entropy)๋ฅผ ์‚ฌ์šฉํ•˜์˜€๊ณ , ๊ฒฝ์‚ฌ ํ•˜๊ฐ•๋ฒ• ์ตœ์ ํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜(Gradient Descent Optimization Algorithm)์œผ๋กœ

Adam(10)์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๋„คํŠธ์›Œํฌ์˜ ๋งˆ์ง€๋ง‰ ๋‘ ๊ฐœ์˜ FC ์ธต์„ ์ œ์™ธํ•œ ๋‚˜๋จธ์ง€ ์ธต๋“ค์€ ์‚ฌ์ „ ํ•™์Šต๋œ(pre-trained) ๋„คํŠธ์›Œํฌ์˜ ๊ตฌ์กฐ๋กœ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์˜ ๋ชจ๋ธ์ด ๊ตฌ์„ฑ๋˜์—ˆ๋‹ค.

3.3 One-Hot ์ง„๋‹จ ๊ฒฐ์ • ๋ชจ๋ธ

3.2 ์ ˆ์—์„œ๋Š” ๋‹จ์ผ OCT ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ๋ผ๋ฒจ์ด ์™„๋ฃŒ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด, 27์žฅ์˜ ์ด๋ฏธ์ง€๋ฅผ ํ•˜๋‚˜์˜ ๋ฐ์ดํ„ฐ ์…‹์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ํ™˜์ž์—๊ฒŒ ์•ˆ๊ตฌ ์งˆํ™˜ ์ง„๋‹จ์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋Š” ๊ฒฐ์ • ๋ชจ๋ธ์„ ์„ค๊ณ„ํ•˜์˜€๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” NN๋ชจ๋ธ๋กœ๋Š” FC, LSTM์™€ CNN์„ ์ ์šฉํ•˜์—ฌ ๊ฐ ๋ชจ๋ธ์˜ ํŠน์ง•๋“ค์„ ํ™œ์šฉํ•˜์—ฌ ๊ตฌํ˜„ํ•˜์˜€๋‹ค. ์„ค๊ณ„๋œ ๋ชจ๋ธ๋“ค์˜ ์ž…๋ ฅ(input) ๋ฐ์ดํ„ฐ์˜ ํ˜•ํƒœ๋Š” ๋ชจ๋‘ ๊ฐ™์œผ๋ฉฐ, ๋ผ๋ฒจ ๋ฐ์ดํ„ฐ ๋˜ํ•œ ๊ฐ™๋‹ค.

FC๋Š” ํ•œ ์ธต์˜ ๋ชจ๋“  ๋‰ด๋Ÿฐ์ด ๋‹ค์Œ ์ธต์˜ ๋ชจ๋“  ๋‰ด๋Ÿฐ๊ณผ ์—ฐ๊ฒฐ๋œ ์ƒํƒœ๋ฅผ ์˜๋ฏธํ•˜๋ฉฐ ์‹ ๊ฒฝ๋ง์˜ ๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ๋„คํŠธ์›Œํฌ์ด๋‹ค. ์ด๋Š” ์ฃผ๋กœ CNN์˜ ํ›„์† ์ธต์œผ๋กœ ์—ฐ๊ฒฐ๋˜์–ด 1์ฐจ์› ๋ฐฐ์—ด์˜ ํ˜•ํƒœ๋กœ ๋ณ€ํ™˜ํ•œ ํ›„, ์ฃผ๋กœ ์ด๋ฏธ์ง€๋ฅผ ๋ถ„๋ฅ˜ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๊ณ„์ธต์ด๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” CNN์„ ๊ธฐ๋ฐ˜ํ•œ ๋ชจ๋ธ์„ ๋ฐ”ํƒ•์œผ๋กœ ์ด๋ฏธ์ง€ ์ฒ˜๋ฆฌ๋ฅผ ํ•˜์˜€๊ธฐ ๋•Œ๋ฌธ์—, ํ›„์† ์ธต์œผ๋กœ FC๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€์žฅ ์ผ๋ฐ˜์ ์ด๋‹ค. ๋”ฐ๋ผ์„œ, ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ FC์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์‹ ๊ฒฝ๋ง์„ ์„ค๊ณ„ํ•˜์˜€๋‹ค.

FC ๋ชจ๋ธ์€ ๋„คํŠธ์›Œํฌ ํฌ๊ธฐ(size)๊ฐ€ 256์ธ FC ์ธต์ด 5๊ฐœ๋กœ ์Œ“์—ฌ ์žˆ๊ณ , ํ•œ ์ธต์—์„œ ๋‹ค์Œ ์ธต์œผ๋กœ ์ •๋ณด๊ฐ€ ์ „๋‹ฌ๋˜๊ธฐ ์ „ ๋ฐฐ์น˜ ์ •๊ทœํ™”์™€ ReLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜๊ฐ€ ์‚ฌ์šฉ๋œ๋‹ค. ๋˜ํ•œ, ์ง€์—ญ์  ์ตœ์  ๋ฌธ์ œ๋ฅผ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด FC ์ธต ์‚ฌ์ด์— ๋“œ๋กญ์•„์›ƒ(dropout)(11)์ด 0.25 ๋น„์œจ๋กœ ์„ค์ •ํ•˜์˜€์œผ๋ฉฐ, ๋งˆ์ง€๋ง‰ ์ธต์€ AMD, DME, Drusen์„ ๋ถ„๋ฅ˜ํ•˜๊ธฐ ์œ„ํ•ด softmax ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๋ณธ ๋ชจ๋ธ์˜ ํ•˜์ดํผ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์—ฌ๋Ÿฌ ๋ฐฉ๋ฒ•์œผ๋กœ ์‹คํ—˜์„ ํ†ตํ•ด ์ตœ์ ์˜ ๋ชจ๋ธ๋กœ ๊ตฌํ˜„ํ•˜์˜€๋‹ค. ๋ณธ ๋ชจ๋ธ์— ์‚ฌ์šฉ๋œ ์†์‹ค ํ•จ์ˆ˜๋Š” ๋‹ค์ค‘ ๋ถ„๋ฅ˜ ์†์‹ค ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์˜€๊ณ , ๊ฒฝ์‚ฌ ํ•˜๊ฐ•๋ฒ• ์ตœ์ ํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ Adam์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ์ด ์ด๋ฃจ์–ด์กŒ๋‹ค.

LSTM์€ RNN(Recurrent Neural Network)๋ฅผ ๊ธฐ๋ฐ˜ํ•œ ๋ชจ๋ธ๋กœ, ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์‹œํ€€์Šค(sequence) ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šตํ•œ๋‹ค๋Š” ์ ์—์„œ ์˜๋ฏธ ์žˆ๋Š” ๋ชจ๋ธ์ด๋‹ค. LSTM์€ ์€๋‹‰์ธต์˜ ๋ฉ”๋ชจ๋ฆฌ ์…€์— ์ž…๋ ฅ ๊ฒŒ์ดํŠธ, ๋ง๊ฐ ๊ฒŒ์ดํŠธ, ์ถœ๋ ฅ ๊ฒŒ์ดํŠธ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ๋ถˆํ•„์š”ํ•œ ์ •๋ณด๋Š” ์ง€์šฐ๊ณ , ๊ธฐ์–ตํ•ด์•ผ ํ•  ์ •๋ณด๋“ค์„ ์„ ํƒํ•˜๊ฒŒ ๋œ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ์กด RNN์˜ ๊ธด ์˜์กด ๊ธฐ๊ฐ„์œผ๋กœ ์ธํ•œ ๋ฌธ์ œ์ ์„ ํ•ด๊ฒฐํ•˜๊ณ , ๊ธด ์‹œํ€€์Šค์˜ ์ž…๋ ฅ์„ ์ฒ˜๋ฆฌํ•˜๋Š”๋ฐ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค. ํ•˜์ง€๋งŒ LSTM์€ ๊ณผ๊ฑฐ ์‹œํ€€์Šค์˜ ์ •๋ณด๋ฅผ ํ†ตํ•ด ํ˜„์žฌ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์„ ๊ฐ€๋Šฅํ•œ ๊ตฌ์กฐ์ด๋‚˜, ์—ญ๋ฐฉํ–ฅ์˜ ์—ฐ๊ฒฐ์€ ์กด์žฌํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์— ํ˜„์‹œ์ ์˜ ์ •๋ณด ์˜ˆ์ธก๋ณด๋‹ค ๋ฏธ๋ž˜ ์‹œ์ ์˜ ์ •๋ณด ์˜ˆ์ธก ์‹œ ํ™œ์šฉํ•  ์ˆ˜ ์—†๋‹ค. ํ•˜์ง€๋งŒ, ์–‘๋ฐฉํ•ญ(Bidirectional)์€ ๋‹จ๋ฐฉํ–ฅ LSTM์˜ ๊ธฐ๋Šฅ์— ์—ญ๋ฐฉํ–ฅ์˜ ๊ธฐ๋Šฅ์„ ์ถ”๊ฐ€ํ•œ ๋ชจ๋ธ๋กœ, ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ์˜ ์ˆœ๋ฐฉํ–ฅ(์‹œ๊ฐ„์ˆœ์„œ)์— ๋”ํ•˜์—ฌ ์—ญ๋ฐฉํ–ฅ(์‹œ๊ฐ„์˜ ์—ญ์ˆœ)์œผ๋กœ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•˜์—ฌ, ์ž…๋ ฅ ๋ฐ์ดํ„ฐ์˜ ํŠน์ •ํ•œ ํŠน์ง•์— ๋Œ€ํ•ด ๋ถ„๋ฅ˜๋‚˜ ์˜ˆ์ธก์ด ๊ฐ€๋Šฅํ•˜๋‹ค(12). ์ถ”๊ฐ€์ ์œผ๋กœ, ์–‘๋ฐฉํ•ญ LSTM์€ ๊ณผ๊ฑฐ๋กœ๋ถ€ํ„ฐ ํ˜„์žฌ์˜ ์˜ˆ์ธก ์ •๋ณด์™€ ๋ฏธ๋ž˜๋กœ๋ถ€ํ„ฐ ํ˜„์žฌ์˜ ์˜ˆ์ธก ์ •๋ณด๋ฅผ ๊ฐ€์ง€๊ณ  ์˜ˆ์ธก์„ ํ•˜๋ฏ€๋กœ ๊ธฐ์กด LSTM๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” LSTM์—์„œ ๋‹ค๋ฃจ๋Š” ์ž…๋ ฅ ์‹œํ€€์Šค ๋ฐ์ดํ„ฐ๋ฅผ ์—ฐ์†์ ์ธ 27์žฅ์˜ OCT ์ด๋ฏธ์ง€๋กœ ๊ฐ„์ฃผํ•˜๊ณ  ์‹ ๊ฒฝ๋ง์„ ์„ค๊ณ„ํ•˜์˜€๋‹ค. ์„ค๊ณ„ํ•œ ์‹ ๊ฒฝ๋ง์€ ๋‘ ๊ฐœ ์ธต์˜ ์–‘๋ฐฉํ–ฅ LSTM๊ณผ ๋‘ ๊ฐœ ์ธต์˜ FC๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋‹ค. ๋„คํŠธ์›Œํฌ ํฌ๊ธฐ๋กœ LSTM์˜ ์ฒซ ๋ฒˆ์งธ ์ธต์€ 128, ๋‘ ๋ฒˆ์งธ ์ธต์€ 256์œผ๋กœ ์„ค์ •ํ•˜์˜€์œผ๋ฉฐ, LSTM๊ณผ ์—ฐ๊ฒฐ๋œ FC์˜ ์ฒซ ๋ฒˆ์งธ ์ธต์€ ์ •๋ณด ์†์‹ค์„ ์ตœ์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด 128๋กœ ์„ค์ •ํ•˜์˜€๊ณ , ๋งˆ์ง€๋ง‰ ์ธต์€ 3์œผ๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. ๋˜ํ•œ, ๊ฐ ์ธต ์‚ฌ์ด์—๋Š” ๋“œ๋กญ์•„์›ƒ์„ 0.3์œผ๋กœ ์„ค์ •ํ•˜๊ณ , ์ตœ์ข… ์ถœ๋ ฅ์ธต์—์„œ AMD, DME์™€ Drusen์„ ๋ถ„๋ฅ˜ํ•˜๊ธฐ ์œ„ํ•ด softmax ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๋ณธ ๋ชจ๋ธ์€ ์†์‹ค ํ•จ์ˆ˜๋กœ์จ ํ‰๊ท  ์ œ๊ณฑ๊ทผ ํ•จ์ˆ˜์™€ ๊ฒฝ์‚ฌ ํ•˜๊ฐ•๋ฒ• ์ตœ์ ํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ Adam์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ์ด ์ด๋ฃจ์–ด์กŒ๋‹ค.

CNN์€ 2์ฐจ์› ๋ฐ์ดํ„ฐ์ธ ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ ๋กœ์ปฌ ๋ถ€๋ถ„์˜ ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด, ์ปจ๋ณผ๋ฃจ์…˜(convolution) ์ธต์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง์ ‘ ํ•™์Šตํ•˜๊ณ  ํŒจํ„ด์„ ์‚ฌ์šฉํ•ด ์ด๋ฏธ์ง€๋ฅผ ๋ถ„๋ฅ˜ํ•œ๋‹ค. ์ž…๋ ฅ์ด ์šฉ์ดํ•˜๊ณ  ์ ์€ ๋งค๊ฐœ๋ณ€์ˆ˜๋กœ ํ•™์Šตํ•œ๋‹ค๋Š” ์žฅ์ ์ด ์žˆ์œผ๋ฉฐ, ๊ทธ ์„ฑ๋Šฅ์€ ๋งค์šฐ ํšจ๊ณผ์ ์ž„์ด ์ž…์ฆ๋˜์—ˆ๋‹ค(13). ๊ทธ๋ž˜์„œ ์š”์ฆ˜ ์˜์ƒ ๋ฐ์ดํ„ฐ ์—ฐ๊ตฌ ๋ถ„์•ผ์—์„œ๋Š” ๋Œ€๋ถ€๋ถ„ CNN ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์˜ ์‹ ๊ฒฝ๋ง์„ ํ™œ์šฉํ•˜๊ณ  ์žˆ๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฐ์ดํ„ฐ ์…‹์„ 2์ฐจ์›์ธ ์ด๋ฏธ์ง€๋กœ ๊ฐ„์ฃผํ•˜๊ณ , ๋ฐ์ดํ„ฐ์˜ ๋ชจ๋“  ๊ฐ’์„ ๊ณ ๋ คํ•˜์—ฌ CNN ๋ชจ๋ธ์— ์ ์šฉํ•œ ์‹ ๊ฒฝ๋ง์„ ์„ค๊ณ„ํ•˜์˜€๋‹ค. CNN์€ ๋‹ค๋ฅธ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๊ธฐ์— ์ตœ๋Œ€ํ•œ ๊ฐ„๋‹จํ•˜๊ฒŒ 4๊ฐœ์˜ ์ธต๊ณผ ๊ฐ ์ธต ์‚ฌ์ด์— ๋“œ๋กญ์•„์›ƒ์„ 0.3์œผ๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. ์ฒซ ๋ฒˆ์งธ ์ธต๊ณผ ์„ธ ๋ฒˆ์งธ ์ธต์˜ ๋„คํŠธ์›Œํฌ ์‚ฌ์ด์ฆˆ๋Š” 64๋กœ ์„ค์ •ํ•˜์˜€์œผ๋ฉฐ, ๊ทธ ์‚ฌ์ด์— ์กด์žฌํ•˜๋Š” ์ธต์˜ ๋„คํŠธ์›Œํฌ ์‚ฌ์ด์ฆˆ๋Š” 256์œผ๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. ๋˜ํ•œ, ์ตœ์ข… ์ธต์„ ์ œ์™ธํ•œ ๋ชจ๋“  ์ธต์—์„œ๋Š” ํ™œ์„ฑํ•จ์ˆ˜๋กœ ReLU๊ฐ€ ์‚ฌ์šฉ๋˜์—ˆ์œผ๋ฉฐ, ์ตœ์ข… ์ธต์€ ํ™œ์„ฑ ํ•จ์ˆ˜๋กœ softmax ๋ถ„๋ฅ˜๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ฒŒ ๋˜๋ฉด์„œ AMD, DME์™€ Drusen์„ ๋ถ„๋ฅ˜ํ•˜๊ฒŒ ๋œ๋‹ค.

4. ์‹คํ—˜ ๋ฐ ๊ฒฐ๊ณผ

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๊ธฐ์กด ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ 2 ๋‹จ๊ณ„๋กœ ์—ฌ๋Ÿฌ ์‹ฌ์ธตํ•™์Šต ๊ธฐ๋ฐ˜์˜ ๋ชจ๋ธ์„ ์ œ์•ˆํ•˜์˜€์œผ๋ฉฐ, ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ต ๋ถ„์„ํ•˜์˜€๋‹ค.

ํ‘œ 1. Phase 1, DenseNet121์˜ ๊ฐœ๋ณ„ ๋ถ„๋ฅ˜์˜ ์„ฑ๋Šฅ

Table 1. Phase 1, Performance of individual classification by DenseNet121

Metrics

Precision

Recall

Accuracy

F1-score

Results

1.00

1.00

99.89

1.00

ํ‘œ 2. Phase 2, ์ง‘๋‹จ ๋ถ„๋ฅ˜์˜ ์„ฑ๋Šฅ ๋น„๊ต

Table 2. Phase 2, Comparison of group classification performance

Algorithm

Precision

Recall

Accuracy

F1-score

FC

0.95

0.96

95.85

0.96

BLSTM

0.93

0.95

93.20

0.94

CNN

0.92

0.99

94.33

0.95

4.1 ๋ฐ์ดํ„ฐ ๋ฐ ์‹คํ—˜ ํ™˜๊ฒฝ

๋ณธ ๋…ผ๋ฌธ์˜ ๋ฐ์ดํ„ฐ๋Š” ์ฐฝ์› ๊ฒฝ์ƒ๊ตญ๋ฆฝ๋Œ€ํ•™๊ต ๋ณ‘์›๊ณผ์˜ ํ˜‘๋ ฅํ•˜์—ฌ ํ™˜์ž ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ IRB ํš๋“ ํ›„, ์•ˆ๊ณผ๋กœ๋ถ€ํ„ฐ AMD, DME, Drusen ํ™˜์ž์˜ OCT ์ด๋ฏธ์ง€๋ฅผ ์—ฐ๊ตฌ์˜ ๋ชฉ์ ์œผ๋กœ ์ˆ˜์ง‘ํ•˜์˜€๋‹ค. ๋ณธ ์—ฐ๊ตฌ์˜ ๋ฐ์ดํ„ฐ ์…‹์€ ์ด 71,496์žฅ์˜ OCT ์ด๋ฏธ์ง€์ด๋ฉฐ, AMD 27,675์žฅ, DME 25,137์žฅ, Drusen 18,684์žฅ์œผ๋กœ ๊ฐ ์ด๋ฏธ์ง€๋Š” ์—ฐ์†์ ์ธ 27์žฅ์˜ ์ด๋ฏธ์ง€๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. ๋˜ํ•œ, ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์˜ DenseNet121 ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๊ธฐ ์œ„ํ•ด Kermany et al.(1)์—์„œ ์ œ๊ณตํ•œ ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ ์…‹์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ ์…‹์€ ์ด 84,484์žฅ์œผ๋กœ, ๊ฐ CNV 37,455์žฅ, DME 11,598์žฅ, Drusen 8,866์žฅ, Normal 26,565์žฅ์œผ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. ์‹ ๊ฒฝ๋ง์€ Google์˜ Tensorflow ๋ฒ„์ „ 2.0์„ ํ†ตํ•ด ๊ตฌํ˜„ํ•˜์˜€์œผ๋ฉฐ ๋‘ ๊ฐœ์˜ NVIDIA GeForce RTX 2080 Ti GPU๋ฅผ ์‚ฌ์šฉํ•˜๋Š” Window10 ์ƒ์—์„œ python์œผ๋กœ ๊ตฌํ˜„ํ•˜์˜€๋‹ค.

4.2 ๋‹จ์ผ OCT ์ด๋ฏธ์ง€ ์ง„๋‹จ ๊ฒฐ์ • ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์‚ฌ์šฉ๋œ ๋ชจ๋ธ์€ DenseNet121 ๋ชจ๋ธ๋กœ, ๊ทธ๋ฆผ 3์˜ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์ด๋‹ค. ํ•™์Šต์€ 100 epoch, ํ•™์Šต๋ฅ (learning rate)์€ 0.001๋กœ ์„ค์ •ํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ๋˜ํ•œ, ํ›ˆ๋ จ ์‹œ 40 epoch๊นŒ์ง€ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ์ด ์ง„ํ–‰๋˜์—ˆ์œผ๋ฉฐ ์ดํ›„ ๋ชจ๋ธ์˜ ๋ณ€๋™(fluctuation)์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ์•ˆ์ •ํ™” ํ›ˆ๋ จ์ด ์ง„ํ–‰๋˜์—ˆ๋‹ค. 100 epoch์„ ๊ธฐ์ค€์œผ๋กœ ๊ทธ ์ด์ƒ ๋ฐ˜๋ณต ํ›ˆ๋ จ์‹œ์ผœ๋„ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์€ ๋ณด์ด์ง€ ์•Š์•˜์Œ์„ ํ™•์ธํ•˜์˜€๋‹ค. ๋ณธ ๋ชจ๋ธ์˜ ํ•™์Šต์„ ์œ„ํ•ด ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ๋Š” Kermany et al.(1)์—์„œ ์ œ๊ณตํ•œ 84,484์žฅ์ด๋ฉฐ, ํ•™์Šต ๋ฐ์ดํ„ฐ 83,484์žฅ๊ณผ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ 1,000์žฅ์œผ๋กœ ๋ถ„๋ฅ˜๋˜์–ด ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ 83,484์žฅ์„ ํ†ตํ•ด ๋ชจ๋ธ์˜ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•˜์˜€์œผ๋ฉฐ 1,000์žฅ์„ ํ†ตํ•ด ํ…Œ์ŠคํŠธ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค.

๊ทธ๋ฆผ. 4. Phase 2์˜ ROC curve

Fig. 4. ROC curve of Phase 2

../../Resources/kiee/KIEE.2021.70.7.1036/fig4.png

๋‹จ์ผ OCT ์ด๋ฏธ์ง€ ํ•™์Šต์— ๋Œ€ํ•ด์„œ ์ง„๋‹จ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ์งˆํ™˜ ๋‹น 250์žฅ์”ฉ ์ด 1,000์žฅ์˜ ๋ฐ์ดํ„ฐ๋กœ ํ…Œ์ŠคํŠธ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ํ‘œ 1์—์„œ ์ •๋ฐ€๋„, ์žฌํ˜„์œจ, ์ •ํ™•๋„ ๋ฐ F1-์ ์ˆ˜๊ฐ€ ์•ˆ๊ตฌ ์งˆํ™˜ ์ž๋™ํ™” ์ง„๋‹จ ์‹œ์Šคํ…œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€๋‹ค. ์„ฑ๋Šฅ ๋น„๊ต๋Š” ๊ฐ ๋ถ„๋ฅ˜์˜ ๊ฒฐ๊ณผ ๊ฐ’์„ ํ‰๊ท ํ•˜์˜€๋‹ค. ๋‹จ์ผ OCT ์ด๋ฏธ์ง€ ์ง„๋‹จ ๊ฒฐ์ • ๋ชจ๋ธ์ธ DenseNet121์€ ์ •๋ฐ€๋„, ์žฌํ˜„์œจ ๊ทธ๋ฆฌ๊ณ  F1-์ ์ˆ˜๋Š” 1.00, ์ •ํ™•๋„๋Š” 99.89๋กœ์„œ ๋งค์šฐ ๋†’์€ ์ •ํ™•๋„๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์—ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ๋†’์€ ์ •ํ™•๋„๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์—ˆ๋˜ ์ด์œ ๋Š” DenseNet121์˜ ๋ชจ๋ธ์€ Densely connection์„ ์ด์šฉํ•˜์—ฌ ๊ฐ ๊ณ„์ธต์˜ ์ถœ๋ ฅ์ธ ํ”ผ์ฒ˜๋งต(feature map)์„ ์ด์–ด์ง€๋Š” ๊ฐ ๊ณ„์ธต์˜ ์ž…๋ ฅ๊ณผ concatenation ์—ฐ์‚ฐ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ํŠน์ง• ์ „ํŒŒ๋ฅผ ๊ฐ•ํ™”ํ•œ๋‹ค๋Š” ์žฅ์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ๋˜ํ•œ ์ด 83,484์˜ ๋งŽ์€ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ์„ ์ˆ˜ํ–‰ํ•˜์˜€๊ธฐ์— ๋†’์€ ์ •ํ™•๋„๋ฅผ ์–ป์—ˆ์„ ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋œ๋‹ค. ๊ทธ๋ฆฌ๊ณ  DenseNet121์˜ AUC(Area under an ROC curve)๋Š” 1.00์ด์—ˆ๋‹ค.

DenseNet121์„ ํ†ตํ•ด์„œ ํ•œ ํ™˜์ž์— ๋Œ€ํ•œ ์—ฌ๋Ÿฌ ์žฅ์˜ ๊ฐ OCT ์ด๋ฏธ์ง€์— ๋Œ€ํ•ด ๋ณ‘์ฆ์— ๋Œ€ํ•œ ํŒ๋… ๊ฒฐ๊ณผ๋ฅผ ํ™•๋ฅ ๊ฐ’์„ ์–ป์–ด์„œ ์ด๋ฅผ ๊ทธ๋ฆผ 3์˜ Phase 1์˜ ์ถœ๋ ฅ์ฒ˜๋Ÿผ ํ™•๋ฅ ๊ฐ’์˜ ๋งคํŠธ๋ฆญ์Šค๋กœ ๋งŒ๋“ค์–ด Phase 2์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜์˜€๋‹ค.

4.3 One-Hot ์ง„๋‹จ ๊ฒฐ์ • ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ

4.2 ์ ˆ์˜ ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ํ†ตํ•ด, ๋ณธ ์žฅ์—์„œ๋Š” ๋ผ๋ฒจ์ด ์™„๋ฃŒ๋œ ๋‹จ์ผ OCT ์ด๋ฏธ์ง€๋ฅผ 27์žฅ์”ฉ ๋ฌถ์–ด ํ•˜๋‚˜์˜ ๋ฐ์ดํ„ฐ ์…‹์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ์ฐฝ์› ๊ฒฝ์ƒ๋Œ€๋ณ‘์›์œผ๋กœ๋ถ€ํ„ฐ ์ˆ˜์ง‘ํ•œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ ์…‹์€ ์ด 71,496์žฅ์ด๋ฉฐ, ๋”ฐ๋ผ์„œ ๋ณธ ์žฅ์—์„œ ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์…‹์€ 2,648๊ฐœ์ด๋‹ค. 2,648๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์…‹ ์ค‘ AMD 1,025๊ฐœ, DME 931๊ฐœ์™€ Drusen 692๊ฐœ๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๋‹ค. ํ•˜์ง€๋งŒ, ๋‹ค์–‘ํ•œ ์‹ฌ์ธตํ•™์Šต ๋ชจ๋ธ๋“ค์„ ํ•™์Šต์‹œํ‚ค๊ธฐ์—๋Š” ๋ฐ์ดํ„ฐ ์…‹์˜ ๋ถ€์กฑ์œผ๋กœ ์ด ๋ฐ์ดํ„ฐ ์…‹์˜ 90%๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์…‹์œผ๋กœ ์„ค์ •ํ•˜๊ณ  ๋‚˜๋จธ์ง€๋Š” ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์…‹์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ ์ˆ˜ํ–‰ํ•œ ์‹คํ—˜์€ FC, LSTM์™€ CNN ๋ชจ๋ธ๋กœ ์ˆ˜ํ–‰ํ•˜์˜€์œผ๋ฉฐ, ๋ชจ๋“  ๋ชจ๋ธ์˜ ๋ฐ์ดํ„ฐ ์…‹์˜ ์„ค์ •์€ ๊ฐ™๊ฒŒ ์„ค์ •ํ•˜์˜€๋‹ค.

FC ๋ชจ๋ธ์€ 1000 epoch, ํ•™์Šต๋ฅ (learning rate)๋Š” 0.00001 ๊ทธ๋ฆฌ๊ณ  ์†์‹ค ํ•จ์ˆ˜๋Š” categorical crossentropy๋กœ ์„ค์ •ํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ํ™˜์ž ์ง„๋‹จ ๊ฒฐ์ • ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ธ FC๋Š” ์ •๋ฐ€๋„ 0.95, ์žฌํ˜„์œจ 0.96, ์ •ํ™•๋„๋Š” 95.85 ๊ทธ๋ฆฌ๊ณ  F1-์ ์ˆ˜๋Š” 0.96์„ ์–ป์—ˆ๋‹ค. ๋˜ํ•œ, ๊ทธ๋ฆผ 4๋ฅผ ํ†ตํ•ด FC์˜ AUC๋Š” 0.99๋ฅผ ์–ป์—ˆ์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ๋ณธ ๋ชจ๋ธ์€ ํ•™์Šต์‹œ epoch์„ 200๊นŒ์ง€๋Š” ํ•™์Šต์ด ๋น ๋ฅด๊ฒŒ ๋˜์—ˆ์œผ๋ฉฐ, 200 ์ด์ƒ๋ถ€ํ„ฐ๋Š” ์ฒœ์ฒœํžˆ ํ•™์Šต๋จ์„ ๋ณผ ์ˆ˜ ์žˆ๋‹ค.

BLSTM ๋ชจ๋ธ์€ 300 epoch, ํ•™์Šต๋ฅ (learning rate)๋Š” 0.001 ๊ทธ๋ฆฌ๊ณ  ์†์‹ค ํ•จ์ˆ˜๋Š” MSE(Mean Square Error)๋กœ ์„ค์ •ํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. BLSTM์€ ์ •๋ฐ€๋„ 0.93, ์žฌํ˜„์œจ 0.95, ์ •ํ™•๋„๋Š” 93.20 ๊ทธ๋ฆฌ๊ณ  F1-์ ์ˆ˜๋Š” 0.94๋ฅผ ์–ป์—ˆ๋‹ค. ๋˜ํ•œ, ๊ทธ๋ฆผ 4๋ฅผ ํ†ตํ•ด BLSTM์˜ AUC๋Š” 0.99๋ฅผ ์–ป์—ˆ์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. FC ๋ชจ๋ธ๊ณผ CNN ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ, BLSTM ๋ชจ๋ธ์€ ํ•™์Šต ์‹œ๊ฐ„์ด 3๋ถ„์—์„œ 5๋ถ„์ด ์ข€ ๋” ๊ฑธ๋ ธ๊ณ  ์ •ํ™•๋„๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์„ฑ๋Šฅ์ด ์•ฝ 1%์ •๋„ ์ฐจ์ด๋‚จ์„ ํ™•์ธํ•˜์˜€๋‹ค.

CNN ๋ชจ๋ธ์€ 1000 epoch, ํ•™์Šต ์†๋„(learning rate)๋Š” 0.0001 ๊ทธ๋ฆฌ๊ณ  ์†์‹ค ํ•จ์ˆ˜์€ MSE(Mean Square Error)๋กœ ์„ค์ •ํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. CNN์€ ์ •๋ฐ€๋„ 0.92, ์žฌํ˜„์œจ 0.99, ์ •ํ™•๋„๋Š” 94.33 ๊ทธ๋ฆฌ๊ณ  F1-์ ์ˆ˜๋Š” 0.95๋ฅผ ์–ป์—ˆ๋‹ค. ๋˜ํ•œ, ๊ทธ๋ฆผ 4๋ฅผ ํ†ตํ•ด CNN์˜ AUC๋Š” 0.95๋ฅผ ์–ป์—ˆ์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค.

๋งˆ์ง€๋ง‰์œผ๋กœ ์‹ ๊ฒฝ๋ง์˜ ์„ฑ๋Šฅ์€ ์‚ฌ์šฉ๋œ ํ•™์Šต ๋ฐ์ดํ„ฐ์™€ ํ•™์Šต ๋ฐ ํ…Œ์ŠคํŠธ ๊ณผ์ •์—์„œ ์„ธ๋ถ€์˜ ์ฐจ์ด๋กœ ์ธํ•ด ์„ฑ๋Šฅ์ด ์กฐ๊ธˆ์”ฉ ์ฐจ์ด๋ฅผ ๋ณด์ผ ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ์„ค๊ณ„๋œ ์‹ฌ์ธตํ•™์Šต ๋ชจ๋ธ๋“ค์˜ ์„ฑ๋Šฅ์„ ์ •ํ™•ํžˆ ๋ถ„์„ํ•˜๊ณ  ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด, ๊ฐ ๋ชจ๋ธ๋“ค์„ 10๋ฒˆ์”ฉ ์žฌํ•™์Šตํ•˜๊ณ  ํ…Œ์ŠคํŠธ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ํ‘œ๋Š” ๊ฐ ๋ชจ๋ธ์„ 10๋ฒˆ์”ฉ ํ…Œ์ŠคํŠธํ•˜์—ฌ ์ •ํ™•๋„์˜ ํ‰๊ท  ๋‚ธ ๊ฒฐ๊ณผ์ด๋ฉฐ, FC ๋ชจ๋ธ์€ 95.89%, BLSTM ๋ชจ๋ธ์€ 93.20%์ด๋ฉฐ CNN 94.33%๋กœ ๋ชจ๋“  ๋ชจ๋ธ์ด 90% ์ด์ƒ ์ •ํ™•๋„์˜ ๊ฒฐ๊ณผ๋ฅผ ๋ƒˆ๋‹ค. ๊ฐ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์—์„œ 1%์—์„œ 2% ์ •๋„์˜ ์ฐจ์ด๊ฐ€ ์žˆ์ง€๋งŒ ์ •ํ™•์„ฑ ๊ธฐ์ค€ FC์˜ ์„ฑ๋Šฅ์ด ๋‹ค๋ฅธ DNN (Deep Neural Network) ๋ชจ๋ธ์— ๋น„ํ•ด ์ข‹์•˜๋‹ค.

์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๋Š” CNN์ด FC์— ๋น„ํ•ด์„œ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค. ํ•˜์ง€๋งŒ ๋ณธ ๋…ผ๋ฌธ์—์„œ ์ง‘๋‹จ ๋ถ„๋ฅ˜์—์„œ ์‚ฌ์šฉํ•˜๋Š” ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋Š” 27 x 4์˜ ๋ฐ์ดํ„ฐ๋กœ์„œ ๋น„๊ต์  ๋‹จ์ˆœํ•˜๋‹ค. ๊ทธ๋ฆฌ๊ณ  CNN์€ FC์— ๋น„ํ•ด ๋” ๋ณต์žกํ•œ ๋‰ด๋Ÿด๋„คํŠธ์›Œํฌ๋ฅผ ๊ตฌ์„ฑํ•œ๋‹ค. DNN (Deep Neural Network)์„ ํ›ˆ๋ จ ์‹œ, FC๋ณด๋‹ค CNN์˜ ์ •ํ™•๋„๊ฐ€ ๋†’๊ฒŒ ๋‚˜์˜ค๋‚˜ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ๋ฅผ ํ…Œ์ŠคํŒ… ํ•  ๊ฒฝ์šฐ CNN์˜ ์ •ํ™•๋„๊ฐ€ FC๋ณด๋‹ค ๋‹ค์†Œ ๋‚ฎ๊ฒŒ ๋‚˜์˜ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ณต์žกํ•œ DNN์„ ์‚ฌ์šฉํ•˜๋ฉด์„œ ๋ฐœ์ƒํ•˜๋Š” ๊ณผ์ ํ•ฉ์ด ๋ฐœ์ƒํ•˜์—ฌ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ FC์˜ ์ •ํ™•๋„๊ฐ€ ๋” ๋†’๊ฒŒ ๋‚˜์˜จ ๊ฒƒ์œผ๋กœ ๋ณด์ธ๋‹ค.

5. ๊ฒฐ ๋ก 

์˜๋ฃŒ ๋ถ„์•ผ์—์„œ๋Š” ์ง„๋‹จ์„ ์œ„ํ•ด ์˜๋ฃŒ ์˜์ƒ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•จ์— ๋”ฐ๋ผ์„œ ์˜๋ฃŒ ์˜์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ์ ์šฉํ•œ ์ปดํ“จํ„ฐ ๋น„์ „(vision) ๋ถ„์•ผ์—์„œ์˜ ์—ฐ๊ตฌ๊ฐ€ ํ™œ๋ฐœํžˆ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋˜ํ•œ OCT ์ด๋ฏธ์ง€๋ผ๋Š” ์˜๋ฃŒ ์˜์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ์ ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์‹ฌ์ธตํ•™์Šต ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ์ž„์ƒ์˜ ์•ˆ๊ณผ ์ „๋ฌธ์˜๋“ค์€ ๋‹จ์ผ OCT ์ด๋ฏธ์ง€๊ฐ€ ์•„๋‹Œ ์—ฌ๋Ÿฌ ์žฅ์˜ OCT ์ด๋ฏธ์ง€๋“ค์„ ์ข…ํ•ฉ์ ์œผ๋กœ ์‚ดํŽด๋ณธ ํ›„, ํ™˜์ž์—๊ฒŒ ์ง„๋‹จ์„ ๋‚ด๋ฆฐ๋‹ค. ์ด์— ๋ณธ ์—ฐ๊ตฌ๋Š” ๊ธฐ์กด์˜ ๋‹ค๋ฅธ ๋…ผ๋ฌธ๊ณผ ๋‹ฌ๋ฆฌ ์—ฌ๋Ÿฌ ์žฅ์˜ ์ด๋ฏธ์ง€๋ฅผ ํ†ตํ•ด ํ•œ ํ™˜์ž์—๊ฒŒ ์ง„๋‹จ์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋Š” ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜์˜€๋‹ค. 1๋‹จ๊ณ„์—์„œ DenseNet121 ๋ชจ๋ธ(8)์„ ํ†ตํ•ด ๋‹จ์ผ OCT ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ ์งˆ๋ณ‘์˜ ์‹œ๊ฐ์  ํŠน์ง•์„ ๋ถ„์„ํ•˜์—ฌ AMD, DME, Drusen๊ณผ Normal์„ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์˜€๋‹ค.

๋˜ํ•œ, 1๋‹จ๊ณ„์˜ ๊ฒฐ๊ณผ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 2๋‹จ๊ณ„์—์„œ๋Š” FC, BLSTM๊ณผ CNN์˜ ๋ชจ๋ธ์„ ํ†ตํ•ด ์—ฌ๋Ÿฌ ์žฅ์˜ OCT ์ด๋ฏธ์ง€๋“ค์˜ ์ •๋ณด๋ฅผ ํ†ตํ•ด ์ข…ํ•ฉ์ ์œผ๋กœ ํŒ๋‹จ ํ›„ ํ™˜์ž์˜ ์งˆํ™˜์„ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ์Œ์„ ํ™•์ธํ•˜์˜€๋‹ค. FC์˜ ์ •ํ™•๋„๋Š” 95.85%์ด๊ณ  ๋‹ค๋ฅธ ๋ชจ๋ธ์˜ ์ •ํ™•๋„์™€ ๋น„๊ตํ•ด์„œ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฌ์Œ์„ ํ™•์ธํ•˜์˜€๋‹ค.

Acknowledgements

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No.NRF-2020R1A2C1014855), Regional Innovation Strategy (RIS) through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(MOE), and the BK21 plus program through the National Research Foundation (NRF) funded by the Ministry of Education of Korea (No.5120201013786).

References

1 
D. S. Kermany, M. Goldbaum, W. Cai, C. C. Valentim, H. Liang, S. L. Baxter, A. McKeown, G. Yang, X. Wu, F. Yan, 2018, Identifying medical diagnoses and treatable diseases by image-based deep learning, Cell, Vol. 172, No. 5, pp. 1122-1131DOI
2 
J. Kugelman, D. Alonso-Caneiro, S. A. Read, J. Hamwood, S. J. Vincent, F. K. Chen, M. J. Collins, 2019, Automatic choroidal segmentation in oct images using supervised deep learning methods, Scientific reports, Vol. 9, No. 1, pp. 1-13DOI
3 
J. Wang, R. Ju, Y. Chen, L. Zhang, J. Hu, Y. Wu, W. Dong, J. Zhong, Z. Yi, 2018, Automated retinopathy of prematurity screening using deep neural networks, EBioMedicine, Vol. 35, pp. 361-368DOI
4 
X. Li, T. Pang, B. Xiong, W. Liu, P. Liang, T. Wang, 2017, Convolutional neural networks based transfer learning for diabetic retinopathy fundus image classi๏ฌcation, in 2017 10th international congress on image and signal processing biomedical engineering and informatics (CISP-BMEI), pp. 1-11DOI
5 
C. S. Lee, D. M. Baughman, A. Y. Lee, 2017, Deep learning is effective for classifying normal versus age-related macular degeneration oct images, Ophthalmology Retina, Vol. 1, No. 4, pp. 322-327DOI
6 
S. Kaymak, A. Serener, 2018, Automated age-related macular degeneration and diabetic macular edema detection on oct images using deep learning, in 2018 IEEE 14th International Conference on Intelligent Computer Communication and Processing (ICCP), pp. 265-269DOI
7 
E. Strickland, 2019, Ibm watson, heal thyself: How ibm overpromised and underdelivered on ai health care, IEEE Spectrum, Vol. 56, No. 4, pp. 24-31DOI
8 
G. Huang, Z. Liu, L. Van Der Maaten, K. Q. Weinberger, 2017, Densely connected convolutional networks, in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4700-4708Google Search
9 
B. Tan, A. Wong, K. Bizheva, 2018, Enhancement of morphological and vascular features in oct images using a modi๏ฌed bayesian residual transform, Biomedical optics express, Vol. 9, No. 5, pp. 2394-2406DOI
10 
D. P. Kingma, J. Ba, 2014, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980Google Search
11 
N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, 2014, Dropout: a simple way to prevent neural networks from over๏ฌtting, The journal of machine learning research, Vol. 15, No. 1, pp. 1929-1958Google Search
12 
M. Schuster, K. Paliwal, 1997, Bidirectional recurrent neural networks, IEEE Transactions on Signal Processing, Vol. 45, No. 11, pp. 2673-2681DOI
13 
P. Y. Simard, D. Steinkraus, J. C. Platt, 2003, Best practices for convolutional neural networks applied to visual document analysis, in Icdar, Vol. 3Google Search

์ €์ž์†Œ๊ฐœ

๊น€์ง€์—ฐ (Jiyeon Kim)
../../Resources/kiee/KIEE.2021.70.7.1036/au1.png

Jiyeon Kim graduated from Gyeongsang National University in 2020.

She is pursuing master degree at the Dept of AI Convergence Engiering(BK21), Gyeongsang National University in 2020.

Her research interests includes Machine Learning, and Medical AI.

ํ•œ์šฉ์„ญ (Yongseop Han)
../../Resources/kiee/KIEE.2021.70.7.1036/au2.png

Yongseop Han graduated from Gyeongsang National University in 2001.

He received Master and Ph.D. degree in the same university in 2005 and 2013, respectively.

He worked as an ophthalmologist at Gyeongsang National University Hospital from 2009, and he joined College of Medicine Gyeongsang National University in 2012 as an associate professor.

His research interest includes Retina, Uvea, and Vitreous.

์ด์›…์„ญ (Woongsup Lee)
../../Resources/kiee/KIEE.2021.70.7.1036/au3.png

Woongsup Lee received a B.S. and a Ph.D. in electrical engineering from KAIST in 2006 and 2012, respectively.

Since 2014, he has been working in the Department of Information and Communication Engineering of Gyeongsang National University, South Korea, where he is now an Associate professor.

๊ฐ•ํƒœ์‹  (Taeseen Kang)
../../Resources/kiee/KIEE.2021.70.7.1036/au4.png

Taeseen Kang graduated from Chungnam National University in 2010.

He received Mater degree in the same university in 2016.

He worked as an ophthalmologist at Gyeongsang National University Changwon Hospital from 2019 as an assistant professor.

His research interest includes Cornea, and Oculoplastics.

์ด์„ฑ์ง„ (Seongjin Lee)
../../Resources/kiee/KIEE.2021.70.7.1036/au5.png

Seongjin Lee graduated from Hanyang University in 2006.

He recieved Master and Ph.D. degree in the same university in 2008 and 2015, respectively.

He worked as postdoc in Storage Center Hanyang University till 2017 and became an assistant research professor there.

He joined Gyeongsang National University in 2017 as an assistant professor.

His research interest includes Operating System, Storage System, System Optimization, Avionics, and Machine Learning.

๊น€๊ฒฝํ›ˆ (Kyong Hoon Kim)
../../Resources/kiee/KIEE.2021.70.7.1036/au6.png

Kyong Hoon Kim received the B.S., M.S., and Ph.D. degrees in computer science and engineering from POSTECH, Pohang, South Korea, in 1998, 2000, and 2005, respectively.

From 2005 to 2007, he was a Post-Doctoral Research Fellow with the CLOUDS Laboratory, Department of Computer Science and Software Engineering, University of Melbourne, Australia.

He is currently a professor in School of Computer Science and Engineering, Kyungpook National University, Daegu, South Korea.

From 2007 to 2019, he was a professor in the Department of Informatics, Gyeongsang National University, Jinju, South Korea.

His current research interests include real-time systems, cloud computing, avionics software, and security.

์ด์˜์„ญ (Yeongseop Lee)
../../Resources/kiee/KIEE.2021.70.7.1036/au7.png

Youngseop Lee graduated from Gyeongsang National University in 2020.

He is pursuing master degree at the Dept of Information Science, Gyeongsang National University in 2020.

His research interests includes Machine Learning, Neural Network, Image Generation, and Image Processing.

๊น€์ง„ํ˜„ (Jin Hyun Kim)
../../Resources/kiee/KIEE.2021.70.7.1036/au8.png

Jin Hyun Kim received the Ph.D. degree from the Department of Computer Science and Engineering, Korea University, Seoul, South Korea, in 2011.

He was a Post-Doctoral Fellow with the KAIST, Daejeon, South Korea, and Aalborg University, Denmark.

He was a Researcher with INRIA/IRISA, Rennes, France.

From 2014 to 2019, he was a Post-Doctoral Fellow with the University of Pennsylvania, Philadelphia, PA, USA.

Since 2019, he has been an assistant professor in Department of Information and Communication Engineering, Gyeongsang University, South Korea.

His current research interests include cyber physical systems, realtime systems, machine learning for medical and vehicular autonomy systems.