Semi-CLMT: A Semi-Supervised Framework for Medical Image Segmentation

In medical image segmentation, traditional fully supervised deep learning methods often encounter challenges in acquiring high-quality annotations due to the significant costs involved. Moreover, many existing semi-supervised methods are difficult to extract structural information from unannotated data efficiently and are prone to learning from unreliable targets in consistency regularization training, which often results in sub-optimal performance. To address these challenges, we propose a novel semi-supervised framework (Semi-CLMT) based on contrastive representation learning and mean teacher-based consistency training, which aims to effectively utilize unannotated medical data to improve the segmentation performance. Our framework first introduces an Uncertainty-guided Weighted Cross-Entropy (U-WCE) loss, enabling the student model to learn from reliable soft pseudo-labels generated by the teacher model under the guidance of the computed 2D uncertainty map. To enhance the model's capability to explicitly learn contextual information about the boundaries of the segmentation targets, we additionally propose an IoU-Similarity based Contrastive (IS-C) loss. Furthermore, we design a Swin Transformer-based encoder-decoder architecture, Trans-AE-Unet, as the backbone for better representation learning by integrating global contextual information and local details. Experiments conducted on three public 2D medical image datasets demonstrate that Semi-CLMT achieves superior segmentation performance compared to state-of-the-art semi-supervised segmentation methods and yields competitive or even better performance in comparison to some recent fully-supervised approaches, despite utilizing a limited amount of labelled data.