ConComp: Constraining and compensating latent representations for semi-supervised medical image segmentation.

Abstract

Semi-supervised learning has significantly advanced medical image segmentation under limited annotation scenarios. However, existing semi-supervised medical image segmentation (SSMIS) methods based on consistency regularization and pseudo-labeling may be affected by internal covariate shift (ICS), which may disturb feature stability and reduce pseudo-label reliability during training. In this work, we investigate how ICS affects feature stability and pseudo-label reliability in SSMIS. Based on this analysis, we propose ConComp, a unified framework that mitigates ICS from a statistical perspective by introducing distribution-aware latent constraints and compensating for latent distribution discrepancies. Specifically, we introduce a Distribution-Aware Feature Constraint (DAFC) mechanism that derives distribution-aware latent constraints from unlabeled feature statistics and residually integrates them into labeled features, thereby reducing changes in their distributions during training. In addition, we propose a Latent Distribution Discrepancy Compensation (L2DC) mechanism that treats the discrepancy between labeled and unlabeled representations as an informative signal to guide feature fusion, thereby compensating for missing structural cues and improving generalization. Extensive experiments on three challenging SSMIS benchmarks, including BraTS2019, LiTS2017, and Pancreas-NIH, demonstrate that ConComp consistently outperforms state-of-the-art methods. The code is publicly available at https://github.com/1jayyang/ConComp.