Mask-Enhanced Edge-Aware Knowledge Distillation for Medical Image Segmentation

Knowledge distillation has become a key paradigm for transferring knowledge from large-scale teacher models to compact student networks, effectively balancing computational efficiency with performance retention. In this paper, we propose a knowledge distillation method based on a mask adaptive mechanism, which efficiently transfers the rich feature information from a high-performance teacher network to a lightweight student network, thereby improving the accuracy of 3D medical image segmentation. Specifically, our method introduces a mask module that generates spatial attention maps using learnable mask tokens, which are weighted via a probabilistic branch to achieve fine-grained feature alignment between the teacher and student in local regions. In addition, to better preserve structural details, we incorporate an edge-aware loss based on the gradient differences along the depth, height, and width directions as a regularization term in the overall loss. To further enhance knowledge transfer, the student features are used to adaptively adjust the mask distribution after a warm-up phase. Experiments on 3D medical image datasets with complex structures, such as WORD and Synapse, demonstrate that our method outperforms others in terms of segmentation accuracy, offering an efficient knowledge distillation strategy that effectively balances performance and computational efficiency.