Episodic Sampling for Class-Imbalanced CT Body Composition Segmentation
A new study from arXiv (2605.20405) proposes episodic sampling, borrowed from few-shot learning, to address class imbalance in medical image segmentation. Unlike loss-based reweighting or standard sampling, episodic sampling explicitly controls which classes appear in each batch, ensuring rare-class exposure. The method was evaluated on body composition segmentation in CT scans using nine muscle and adipose tissue classes from 210 scans of the public SAROS dataset. Comparisons against random and weighted sampling were performed under full- and low-data regimes.
Key facts
- Episodic sampling from few-shot learning is adapted for fully supervised segmentation.
- Class imbalance is tackled by promoting class-balanced batch construction.
- Method decoupled from conventional metric-learning context.
- Evaluated on nine muscle and adipose tissues from 210 SAROS CT scans.
- Compared against random and weighted sampling.
- Training performed under full- and low-data regimes.
- Loss-based and sampling approaches do not explicitly control class appearance in batches.
- Study focuses on body composition segmentation.
Entities
—