Episodic Sampling Boosts Medical Image Segmentation with 12x More Iterations Before Overfitting

Class imbalance is a persistent problem in medical image segmentation, where rare tissues are often underrepresented during training. Existing approaches use loss reweighting or sampling strategies, but neither explicitly controls which classes appear in each batch. Iason Skylitsis, Dimitrios Karkalousos, and Ivana Išgum from the University of Amsterdam propose borrowing episodic sampling from few-shot learning to construct class-balanced batches in a fully supervised setting. They evaluate on nine muscle and adipose tissues from 210 CT scans of the public SAROS dataset.

Under full-data training, all three sampling strategies performed comparably (mean Dice ~0.88). However, in low-data regimes, episodic sampling achieved 0.787 Dice versus 0.758 (random) and 0.762 (weighted). Crucially, when matched for training iteration budgets, random and weighted sampling overfitted earlier, while episodic sampling continued to improve for roughly three times more iterations before plateauing. This reveals the iteration budget as an under-recognized confound in comparing sampling strategies.

The residual advantage of episodic sampling is attributed to an implicit regularization effect from class-balanced batches. The method is model-agnostic and low-cost, requiring no architectural changes. The authors provide open-source code and advocate for iteration-aware evaluation protocols, especially for small medical imaging datasets.