CMRVAE: Contrastive margin-restrained variational auto-encoder for class-separated domain adaptation in cardiac segmentation

Abstract

Unsupervised Domain Adaptation (UDA) is a promising strategy for representing unlabeled data through domain alignment. Nonetheless, a considerable number of whole-domain alignment techniques often neglect the essential interconnections between pixels and patches across distinct domains that exhibit analogous semantic characteristics. This oversight can hinder their ability to manage semantic variations across domains and to create a discriminative embedding for different classes, ultimately leading to reduced discrimination and poor generalization. This paper presents a novel UDA method for medical image analysis, termed CMRVAE. The proposed method is composed of a margin-restrained variational auto-encoder (MR-VAE) and a class-separation patch-level manifold clustering (CPMC) module. The MR-VAE embeds an adaptive margin-based enhancement technique through an innovative variational inference for optimal encoder mapping in UDA. The CPMC module integrates multi-granularity class information into the manifold for improved preparatory work before UDA. Experimental results on three cardiac datasets show that the proposed method achieves substantially enhanced accuracy compared to the state-of-the-art unsupervised approaches.