Dual-Contrastive Dual-Consistency Dual-Transformer: A Semi-Supervised Approach to Medical Image Segmentation

Medical image segmentation serves as a crucial under-pinning for a myriad of clinical applications. The advent of deep learning techniques has significantly propelled advancements in this field. However, challenges persist due to the limited availability of labelled medical imaging data and the substantial cost of data annotation. This paper introduces a novel semi-supervised learning strategy, amalgamating pseudo-labelling and contrastive learning with a consistency regularization framework. This innovative approach incorporates a modified contrastive learning strategy and a confidence-aware pseudo-labeling strategy, both of which are integrated into a dual-segmentation network ensemble learning structure. Inspired by the recent success of self-attention mechanisms, we harness the power of the Vision Transofmer(ViT) within our proposed semi-supervised framework, and conduct a comprehensive comparison among various combinations of ViT and Convolutional Neural Network(CNN) with the proposed strategy. The efficacy of our proposed method is validated using a publicly available medical image segmentation dataset, where it demonstrates state-of-the-art performance against established methods. The proposed method, all baseline methods, and dataset are available at https://github.com/ziyangwang007/CV-SSL-MIS.