Oct 25, 2024

KoSAIM

Abstract

OSTEO

Enhanced Multi-Class Classification on Osteoporosis Severity in Chest X-Rays Using Knowledge Distillation and Proxy Labels

Junhyeok Park

Object 

๋ณธ ์—ฐ๊ตฌ์˜ ๋ชฉ์ ์€ ํ‰๋ถ€ X์„  ์˜์ƒ(Chest X-Ray, CXR)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ •์ƒ, ๊ณจ๋‹ค๊ณต์ฆ ๋ฐ ๊ณจ๊ฐ์†Œ์ฆ์„ ๋ถ„๋ณ„ํ•˜๊ธฐ ์œ„ํ•œ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒํ•˜๊ณ , ๋‹ค์–‘ํ•œ ์™ธ๋ถ€ ๊ฒ€์ฆ์„ ํ†ตํ•ด ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉ์ ์œผ๋กœ ํ•œ๋‹ค.

Methods

Figure 1: Teacher-Student ๊ตฌ์กฐ๋ฅผ ํ†ตํ•œ ํ•™์Šต๊ณผ Proxy-Labeled Data ์ ์šฉ ๊ณผ์ •

 ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” CXR์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ณจ๋‹ค๊ณต์ฆ, ๊ณจ๊ฐ์†Œ์ฆ, ์ •์ƒ์„ ๋ถ„๋ฅ˜ํ•˜๋Š” ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•ด Teacher-Student ๊ตฌ์กฐ์˜ ์ง€์‹ ์ฆ๋ฅ˜ ๋ฐฉ๋ฒ•์„ ํ™œ์šฉํ•œ๋‹ค[1]. ๋˜ํ•œ, ์ค€์ง€๋„ ํ•™์Šต ๋ฐฉ๋ฒ• ์ค‘ ํ•˜๋‚˜์ธ Proxy-label Method๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ๊ฐ•๊ฑดํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ๊ฒƒ์„ ๋ชฉ์ ์œผ๋กœ ํ•œ๋‹ค[2]. 

 Teacher-Student ๊ตฌ์กฐ์—์„œ ํ•ด๋‹น Teacher ๋ชจ๋ธ์€ ์ž…๋ ฅ ์˜์ƒ์— ์•ฝํ•œ ์ฆ๊ฐ•์„, Student ๋ชจ๋ธ์€ ๊ฐ•ํ•œ ์ฆ๊ฐ•์„ ์ ์šฉํ•œ ์ƒํƒœ๋กœ ํ•™์Šตํ•˜๋ฉฐ[3] Offline ๊ตฌ์กฐ์ธ Teacher ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ์ง€์ˆ˜์ด๋™ํ‰๊ท (Exponential Moving Average)์„ ํ†ตํ•ด ๋ฐ˜๋ณต์ ์œผ๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ „๋‹ฌ๋ฐ›์•„ Student ๋ชจ๋ธ์ด ์›ํ™œํ•˜๊ฒŒ ํ•™์Šต๋  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” Knowledge Distillation(KD) ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•œ๋‹ค.[4] ํ•ด๋‹น ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ํ™˜์ž์˜ ํŠน์ • ๋ถ€์œ„์˜ ์ƒํƒœ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ „์ฒด์ ์ธ ์ƒํƒœ๋ฅผ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์ดฌ์˜๊ธฐ๊ธฐ์—์„œ ํš๋“ํ•œ ๋ฐ์ดํ„ฐ์—๋„ ๊ฐ•๊ฑด์„ฑ์„ ๊ฐ€์งˆ ์ˆ˜ ์žˆ๊ฒŒ ๋œ๋‹ค.

 ๋˜ํ•œ ํ‰๋ถ€X์„  ์˜์ƒ๊ณผ DEXA(Dual-Energy X-ray Absorptiometry; ์ด์ค‘์—๋„ˆ์ง€ X์„  ํก์ˆ˜๊ณ„์ธก๋ฒ•) ๊ฒ€์ง„ ๊ฒฐ๊ณผ๊ฐ€ ์ง์„ ์ด๋ฃจ์–ด์•ผ ํ•˜๋Š” ๋ฐ์ดํ„ฐ์˜ ํŠน์„ฑ์ƒ ๋‹ค๊ธฐ๊ด€ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์ด ์–ด๋ ค์šด ์ ์„ ๊ณ ๋ คํ•˜์—ฌ, ์ค€์ง€๋„ ํ•™์Šต์˜ Proxy-label Method ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด Public Dataset์ธ CheXpert[5]๋ฅผ ์ถ”๊ฐ€๋กœ ํ™œ์šฉํ•˜์˜€๋‹ค.(Figure 1. ์ฐธ๊ณ ) ํ•ด๋‹น ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ๊ฐ•๊ฑดํ•จ์„ ํ–ฅ์ƒ์‹œ์ผฐ๊ณ , ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ์…‹์„ ํ†ตํ•ด ๋ชจ๋ธ์˜ ํ–ฅ์ƒ๋œ ๊ฐ•๊ฑดํ•จ์„ ๊ฒ€์ฆํ•˜์˜€๋‹ค.

Results 

๋ณ‘์› A์˜ ํด๋ž˜์Šค๋ณ„ ๋ฐ์ดํ„ฐ ์ˆ˜๋ฅผ ๋งž์ถ˜ ๋‚ด๋ถ€ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ(16.1% ๋‚จ์„ฑ; ํ‰๊ท  ์—ฐ๋ น 58.7 ยฑ 6.76)์™€ ๋™์ผํ•œ ๋ณ‘์› A์˜ ์™ธ๋ž˜, ์ž…์›, ์‘๊ธ‰์‹ค, ๊ฑด๊ฐ•๊ฒ€์ง„ ๋ฐ ์ฆ์ง„์„ผํ„ฐ์—์„œ ๋‹ค๋ฅธ ๊ธฐ๊ฐ„์— ์ˆ˜์ง‘๋œ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ(11.1% ๋‚จ์„ฑ; ํ‰๊ท  ์—ฐ๋ น, 59.01 ยฑ 6.65), 2์ฐจ ์ข…ํ•ฉ ๋ณ‘์› B(44.5% ๋‚จ์„ฑ; ํ‰๊ท  ์—ฐ๋ น, 59.38 ยฑ 7.31) ๋ฐ์ดํ„ฐ, ๋ณดํ›ˆ ๋ณ‘์› C(56.2% ๋‚จ์„ฑ; ํ‰๊ท  ์—ฐ๋ น, 73.64 ยฑ 6.74) ๋ฐ์ดํ„ฐ ๋“ฑ์˜ ๊ฒ€์ฆ ๊ฒฐ๊ณผ์—์„œ 3 Class์— ๋Œ€ํ•œ ์ •ํ™•๋„๊ฐ€ ๊ฐ๊ฐ 0.730, 0.732, 0.769, 0.668 ์œผ๋กœ ๊ฒ€์ฆ๋˜์—ˆ๋‹ค.(Table 1. ์ฐธ๊ณ )

Table 1: ๊ณจ๋‹ค๊ณต์ฆ ์ •๋„์˜ Ablation Study ๊ฒฐ๊ณผ


Hospital A (Internal)

Hospital A (External)

Hospital B

Hospital C

Baseline

0.531

0.641

0.619

0.541

+ KD

0.711

0.693

0.752

0.633

+ Proxy Label (Ours)

0.730

0.732

0.769

0.668

Note:- KD, Knowledge Distillation

Conclusions 

์ง€์‹ ์ฆ๋ฅ˜ ๋ฐฉ๋ฒ•๊ณผ Proxy-label์„ ํ™œ์šฉํ•œ ์ค€์ง€๋„ ํ•™์Šต์„ ํ†ตํ•ด ๊ณจ๊ฐ์†Œ์ฆ๊ณผ ๊ณจ๋‹ค๊ณต์ฆ์˜ CXR ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๋ฅผ ์œ„ํ•œ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ํ†ตํ•ด์„œ, ๋‹ค์–‘ํ•œ ์™ธ๋ถ€ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์—์„œ๋„ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋‚˜ํƒ€๋ƒˆ๊ณ  ์ด๋ฅผ ํ†ตํ•ด ์ ‘๊ทผ์„ฑ์ด ์ œํ•œ์ ์ธ DEXA ๊ฒ€์‚ฌ์™€ ๋‹ค๋ฅด๊ฒŒ CXR์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณจ๊ฐ์†Œ์ฆ ๋ฐ ๊ณจ๋‹ค๊ณต์ฆ ์Šคํฌ๋ฆฌ๋‹์— ์ด์šฉ๋  ์ˆ˜ ์žˆ์Œ์„ ํ™•์ธํ•˜์˜€๋‹ค.

 

Reference

[1] Tarvainen, Antti, and Harri Valpola. "Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results." Advances in neural information processing systems 30 (2017).

[2] Lee, Dong-Hyun. "Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks." Workshop on challenges in representation learning, ICML. Vol. 3. No. 2. 2013.

[3] Buslaev, Alexander, et al. "Albumentations: fast and flexible image augmentations." Information 11.2 (2020): 125.

[4] Li, Lujun, and Zhe Jin. "Shadow knowledge distillation: Bridging offline and online knowledge transfer." Advances in Neural Information Processing Systems 35 (2022): 635-649.

[5] Irvin, Jeremy, et al. "Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison." Proceedings of the AAAI conference on artificial intelligence. Vol. 33. No. 01. 2019.

* Keywords

Chest Radiograph, Bone Mass Density, Osteopenia

PROMEDIUS INC.

Copyright 2025 PROMEDIUS INC. All rights reserved.

13, Olympic-ro 35da-gil, Songpa-gu, Seoul, 05510 Republic of Korea

PROMEDIUS INC.

Copyright 2025 PROMEDIUS INC. All rights reserved.

13, Olympic-ro 35da-gil, Songpa-gu, Seoul, 05510 Republic of Korea