Semi-Supervised Volumetric Medical Image Segmentation via Class Prototype Guided Distribution-Aligned Representation Learning

Abstract

We present SemiCRL, a novel framework for volumetric medical image segmentation that formulates an innovative contrastive learning methodology in a semi-supervised learning setting. We leverage the pseudo-labels generated in semi-supervised learning to guide the selection of negative samples for our contrastive learning, aiming to alleviate the class collision issue and learn enhanced class-discriminative latent representations. However, to address the inaccuracies in pseudo-labels, which stem from the empirical distribution misalignment between labeled and unlabeled data, we introduce a pseudo-label refinement strategy based on class prototypes computed from learned latent representations. Furthermore, our contrastive learning utilizes class prototypes as powerful reference points to enforce the alignment of latent-space distribution of labeled and unlabeled data, thus fostering knowledge transfer from labeled to unlabeled data, which in turn enhances the generation of accurate pseudo-labels in semi-supervised learning. Experiments on two public medical image datasets demonstrate our proposed method outperforms existing state-of-the-art semi-supervised approaches.