Contrmix: Progressive Mixed Contrastive Learning for Semi-Supervised Medical Image Segmentation

Abstract

While medical image segmentation has achieved impressive progress, it usually being constrained by labor-intensive and costly pixel-wise annotations. The existing semi-supervised learning methods ignore the inherent imbalance and high similarity of different categories in medical images. To address the above issues, we present a Progressive Mixed Contrastive Learning (ContrMix) framework, which contains a Cycle-mix module and a mix-based Contrastive Learning module. In Cycle-mix, a progressive mixing strategy with a cycle loss is designed to enforce the consistency between the mixed segmentation and corresponding generated mixing samples, effectively enhancing the ability to learn geometric features of the imbalanced medical data. We also introduce a mix-based Contrastive Learning module that learns the inter-instance similarities between the mixed patches and the original ones, which encourages the model to learn background-invariant representations from samples under different distortions and improves the semantic discrimination of high similarity categories. We conduct extensive experiments on the ACDC dataset and LA dataset and our method outperforms other state-of-the-art semi-supervised approaches.