Balanced feature fusion collaborative training for semi-supervised medical image segmentation

Collaborative learning is a fundamental component of consistency learning. It has been extensively utilized in semi-supervised medical image segmentation, primarily based on the learning of multiple models from each other. However, existing semi-supervised collaborative image segmentation methods face two primary issues. Firstly, these methods fail to fully leverage the hidden knowledge within the models during a knowledge exchange, resulting in inefficient knowledge sharing and limited generalization capabilities. To address this, we propose a novel approach, termed ‘fusion teacher’, which merges the knowledge of two models at the feature-level. This enhances the efficiency of knowledge exchange between models and generates more accurate pseudo-labels for consistency learning. Secondly, the initial and intermediate stages of collaborative learning are hindered by a significant performance gap between the fusion teacher and student models, impairs effective knowledge transfer. Our approach advocates a gradual increase in the dropout rate. This strategy enhances the transfer efficiency of knowledge from a fusion teacher to a student model. To demonstrate the efficacy of our method, we conduct experiments on the ISIC, ACDC, and AbdomenCT-1K datasets. Our approach achieves Dice scores of 87.4%, 84.8%, and 84.5%, respectively, with 10% labelled data. Compared with the current state-of-the-art (SOTA) methods, our method demonstrates strong competitiveness. •A new collaborative semi-supervised learning framework is designed.•Better pseudo labels are obtained by the teacher model through feature fusion.•A method for balancing the learning speed of teacher and student models is proposed.•Our method has strong competitiveness compared with the state-of-the-art method on three popular datasets.