Disentangling Sampling from Training Budget in Class-Imbalanced CT Body Composition Segmentation

Class imbalance is a fundamental challenge in medical image segmentation, where frequent classes typically dominate training at the expense of rare classes. Loss-based approaches mitigate imbalance by reweighting the per-pixel loss within the batch, while sampling strategies control which images enter the batch. Yet neither explicitly controls which classes appear within the batch, leaving rare-class exposure only partially rebalanced. In this work, we adopt episodic sampling from few-shot learning to promote class-balanced batch construction in a fully supervised setting. We decouple episodic sampling from its conventional metric-learning context and evaluate it in body composition segmentation in CT. We compare episodic sampling against random and weighted sampling on nine muscle and adipose tissues, derived from 210 scans of the public SAROS dataset. Training is performed under full- and low-data regimes, with additional comparisons under matched training iteration budgets. Under full-data training, all three strategies performed comparably (mean Dice 0.882 for episodic, 0.878 for random and weighted). Under low-data training, episodic sampling outperformed random and weighted (0.787 vs. 0.758 and 0.762), driven by a 12-fold difference in training iterations. Under matched training budgets, random and weighted overfit earlier, while episodic improved for approximately three times more iterations before plateauing. Our findings identify the training iteration budget as under-recognized confound in sampling strategies, motivating iteration-aware evaluation protocols for small datasets. Furthermore, the residual advantage of episodic sampling is consistent with an implicit regularization effect of class-balanced batches, offering a low-cost, model-agnostic strategy for class-imbalanced medical image segmentation. Code is available at https://github.com/iasonsky/episodic-sampling.

한 줄 요약

**[CT 신체 조성 분할 / Episodic Sampling]** Episodic sampling이 SAROS muscle·adipose 9 조직 분할에서 low-data regime(Dice 0.787 vs 0.758)에서 random·weighted 능가 — 12× 더 많은 iteration 효과, matched budget에서도 plateau가 3× 늦어 implicit regularization 효과 확인.

핵심 기여도

Medical image segmentation의 fundamental challenge로 class imbalance를 진단 — 빈번 클래스가 학습을 dominate하며 rare 클래스 희생.
Loss-based 접근이 batch 내 per-pixel loss reweighting으로 imbalance 완화, sampling 전략이 batch에 들어가는 이미지 제어. 그러나 둘 다 batch 내 어떤 클래스가 등장할지를 explicitly 제어하지 못해 rare-class exposure가 partially만 rebalance됨을 지적.
Few-shot learning의 episodic sampling을 채택 — fully supervised 환경의 class-balanced batch 구성 촉진. Conventional metric-learning context에서 분리해 CT body composition segmentation에 평가.
CT 9 근육·지방 조직(SAROS 데이터셋 210 scan 유래)에서 episodic을 random·weighted sampling과 비교 — full data·low data regime 학습 수행, matched training iteration budget에서 추가 비교 동반.

핵심 아이디어

Class-imbalanced segmentation에서 loss reweighting·sampling 어느 것도 batch 내 class composition을 명시 제어하지 못하므로, few-shot learning의 episodic sampling을 supervised로 가져와 class-balanced batch를 강제하면 low-data regime에서 implicit regularization 효과로 plateau를 늦추고 정확도를 향상시킬 수 있다 — training iteration budget은 sampling 평가의 under-recognized confounder.

기술적 접근법

**방법론**: Episodic sampling을 fully supervised CT 분할에 적용.
**핵심 기법**: (1) Few-shot의 episodic sampling을 metric-learning context에서 decouple, (2) Batch 내 class-balanced 구성을 강제, (3) SAROS 데이터셋 210 scan에서 9 muscle·adipose 조직 분할, (4) Random·weighted sampling과 full data·low data regime 비교, (5) Matched training iteration budget 비교로 iteration confounder 검출.

주요 결과

Full data: 3 전략 모두 비슷한 성능 — episodic 0.882, random·weighted 0.878.
Low data: episodic 0.787 vs random 0.758·weighted 0.762 — 12× iteration 차이가 주된 동인.
Matched budget: random·weighted가 더 일찍 overfit, episodic이 약 3× 더 오래 향상 후 plateau.
Training iteration budget이 sampling 평가의 under-recognized confounder임을 식별.
Code: github.com/iasonsky/episodic-sampling.

의의 및 한계

**의의**: Sampling 전략 평가 protocol에 iteration-aware 시각 도입 — 기존 비교의 confounder 노출, episodic sampling의 implicit regularization 효과 확인으로 low-cost·model-agnostic class-imbalance 처리 도구 제공. **한계**: SAROS 9 조직 단일 task로 다른 medical imaging task 일반화 추가 검증, episodic batch 구성의 hyperparameter 부담, full data에서는 명확한 우위 부재.

실용적 활용

Class-imbalanced 의료 영상 분할.
Sampling 전략 평가 protocol의 reference.
Few-shot·imbalanced learning 일반 도구.