Attention correction feature and boundary constraint knowledge distillation for efficient 3D medical image segmentation

Xiangchun Yu, Longxiang Teng, Dingwen Zhang, Jian Zheng, Hechang Chen

Abstract: Highlights•We observe residual connections and attentions function as corrections based on NMF.•We conduct feature-based KD without incorporating auxiliary modules into the teacher.•We introduce a learnable attention correction factor to bridge the knowledge gap.•We develop the BCD to improve students’ perception of boundaries.•Experimental results show that our proposed ACF-BCKD achieves SOTA performance.