Reproduce “Semantic Class Distribution Learning for Debiasing Semi-Supervised Medical Image Segmentation” on OpenResearch

About this paper

Yingxue Su, Yiheng Zhong, Keying Zhu, Zimu Zhang, Zhuoru Zhang, Yifang Wang, Yuxin Zhang, Xinyuan Zheng and 2 more

The Semantic Class Distribution Learning (SCDL) framework addresses class imbalance in semi-supervised medical image segmentation by explicitly learning and aligning class-conditional feature distributions. It improves segmentation accuracy, particularly for small, under-represented anatomical structures, achieving consistent gains across multiple baselines and up to 10.67 percentage points increase in mean Dice Similarity Coefficient on challenging datasets.

Key insights

• Explicitly modeling and aligning class-conditional feature distributions within the embedding space provides a more fundamental solution to class imbalance than output- or loss-level adjustments.
• The combination of bidirectional alignment for feature organization and direct semantic anchoring from labeled data is crucial for achieving both discriminative feature learning and robust geometric boundary quality for all classes.
• Leveraging limited labeled data to create semantic anchors effectively stabilizes and refines minority-class proxies, enabling more reliable segmentation even with sparse annotations.

Summary by alphaXiv.