Multi-Level Contrastive Student-Teacher Structure for Semi-Supervised Medical Image Segmentation

The development of deep learning models is constrained by the scarcity of annotated data in medical image analysis. To address this challenge, semi-supervised and contrastive learning have shown significant potential, particularly in reducing the reliance on precise pixel-level annotations. In this study, we combine contrastive learning with the semi-supervised framework to propose a novel Multi-level Contrastive Mean Teacher framework (MCMT) to improve the discriminative features understanding and extraction of the segmentation model in similar samples. Specifically, the MCMT framework employs the student-teacher structure, in addition to establishing consistency between the two models and establishing feature contrasts by the different stages outputs of the model encoder to optimize the segmentation model through consistency and contrastive loss. Experiments on two public segmentation tasks show that by incorporating contrastive learning, the MCMT framework outperforms the traditional method at feature representation and extraction, achieving better segmentation accuracy. It demonstrates that integrating contrastive learning into the semi-supervised framework could effectively improve feature representation and reduce the dependency of the segmentation model on annotated data.