Dual cross-attentive mutual teaching for semi-supervised 3D medical segmentation

Abstract

Semi-supervised learning can reduce the dependence on large-scale labeled data in 3D medical image segmentation.In this work, we propose a new Dual Crossed Attention Mutual Teaching (DCA-MT) framework that effectively utilizes both labeled and unlabeled data by integrating high-dimensional feature alignment, semantic-level crossed attention, and bidirectional knowledge distillation. Specifically, we employ a two-branch VNet architecture where the teacher-student network co-evolves through mutual mentoring and collaborative learning.To enhance representation consistency, we introduce maximum mean difference (MMD) loss and inter-class and intra-class contrast constraints to achieve global feature distribution alignment and class-level separability. A multi-head cross-attention module is designed to facilitate fine-grained semantic interaction between the two networks, allowing the two branches to dynamically exchange complementary features.In addition, the two-way mutual distillation strategy ensures that teacher and student networks benefit from each other's knowledge. Numerous experiments on the left atrial and pancreatic nih datasets show that our proposed approach has better performance and verifies the effectiveness and robustness of DCA-MT.