Edge-Guided and Feature-Aligned Coordinated Knowledge Distillation for Polyp Segmentation

Accurate polyp segmentation is crucial for the early screening of colorectal cancer. However, conventional Knowledge Distillation (KD) methods often inadequately address challenges such as blurred edges and semantic gaps in polyp segmentation. To overcome these limitations, this paper proposes a novel coordinated knowledge distillation framework. This framework aims to efficiently transfer knowledge from a large teacher network to a lightweight student network, significantly enhancing its polyp segmentation capabilities. It strengthens edge perception through edge-constrained distillation focusing on boundary pixels, bridges the semantic gap using a feature alignment mechanism that reorganizes channel and spatial information, and is further supplemented by an attention-guided approach to optimize the student network's focus on critical spatial regions. Extensive experiments on multiple public polyp segmentation datasets demonstrate that our framework enables lightweight student models to achieve substantial improvements over baseline student models trained only with segmentation loss. For instance, on the Kvasir dataset, our method boosts the student model's mDice from 0.893 to 0.914 and mIoU from 0.835 to 0.864. Similarly, on the CVC-ClinicDB dataset, mDice increased from 0.895 to 0.933 and mIoU from 0.833 to 0.882. This level of performance also significantly surpasses conventional KD and various other comparative methods. Furthermore, the enhanced student models achieve performance comparable to the large teacher network, while markedly improving edge segmentation accuracy and the overall quality of polyp region identification. This research provides an effective and lightweight solution, promising to enhance polyp segmentation for real-time clinical diagnosis.