EA-Net: Edge Attention Network for Brain Tumour Segmentation in MRI.
Accurate brain tumour segmentation is crucial for clinical diagnosis and treatment planning, yet remains challenging due to the scale diversity of tumour regions, ambiguous boundary structures, and highly irregular shapes.
We propose a novel Edge Attention Network that integrates two key components: a Multi-Scale Context Fusion Module to dynamically adjust receptive fields and capture multi-scale contextual information, and an Edge Segmentation Module that explicitly extracts tumour boundaries and injects them into the backbone as spatial attention to refine segmentation details, particularly at edges.
Experiments show that our model achieves Dice coefficients of 90.37% for Tumour Core (TC) and 88.91% for Whole Tumour (WT) on the BraTS2021 dataset. In cross-dataset generalisation tests on BTM-PVS, it maintains strong performance with 75.20% TC and 74.20% WT.
The proposed method demonstrates superior segmentation accuracy and robust generalisation capability, highlighting its clinical potential and offering new insights for medical image segmentation.
We propose a novel Edge Attention Network that integrates two key components: a Multi-Scale Context Fusion Module to dynamically adjust receptive fields and capture multi-scale contextual information, and an Edge Segmentation Module that explicitly extracts tumour boundaries and injects them into the backbone as spatial attention to refine segmentation details, particularly at edges.
Experiments show that our model achieves Dice coefficients of 90.37% for Tumour Core (TC) and 88.91% for Whole Tumour (WT) on the BraTS2021 dataset. In cross-dataset generalisation tests on BTM-PVS, it maintains strong performance with 75.20% TC and 74.20% WT.
The proposed method demonstrates superior segmentation accuracy and robust generalisation capability, highlighting its clinical potential and offering new insights for medical image segmentation.
Authors
Fansheng Fansheng, Xingguang Xingguang, Xinya Xinya, Fengxinyun Fengxinyun, Jiexi Jiexi, Xujia Xujia, Yu Yu, Yuanli Yuanli, Changsheng Changsheng
View on Pubmed