Semantic Class Distribution Learning for Debiasing Semi-Supervised Medical Image Segmentation
This work is significant for medical practitioners and researchers working with medical image segmentation, as it tackles the common problem of class imbalance that hinders reliable segmentation of critical minority structures.
This paper addresses the problem of severe class imbalance in semi-supervised medical image segmentation, which causes minority structures to be overwhelmed. The proposed SCDL framework significantly improves segmentation performance on both overall and class-level metrics, achieving state-of-the-art results, especially for minority classes on Synapse and AMOS datasets.
Medical image segmentation is critical for computer-aided diagnosis. However, dense pixel-level annotation is time-consuming and expensive, and medical datasets often exhibit severe class imbalance. Such imbalance causes minority structures to be overwhelmed by dominant classes in feature representations, hindering the learning of discriminative features and making reliable segmentation particularly challenging. To address this, we propose the Semantic Class Distribution Learning (SCDL) framework, a plug-and-play module that mitigates supervision and representation biases by learning structured class-conditional feature distributions. SCDL integrates Class Distribution Bidirectional Alignment (CDBA) to align embeddings with learnable class proxies and leverages Semantic Anchor Constraints (SAC) to guide proxies using labeled data. Experiments on the Synapse and AMOS datasets demonstrate that SCDL significantly improves segmentation performance across both overall and class-level metrics, with particularly strong gains on minority classes, achieving state-of-the-art results. Our code is released at https://github.com/Zyh55555/SCDL.