Abstract

Info

Title:: Brain white matter lesion segmentation using an enhanced U-Net integrated with attention mechanisms

Author(s):: LIU Qingchen1; HAN Xiaoxin1; KUANG Jingfan1; HU Yuchen1; WANG Jianlin2; 1. College of Medical Information Engineering, Gansu University of Chinese Medicine, Lanzhou 730000, China 2. Information Center, the First Hospital of Lanzhou University, Lanzhou 730000, China

Keywords:: Keywords:?mage segmentation brain white matter lesion multi-scale feature calibration attention mechanism

Abstract:: Abstract: Given that current brain white matter lesion segmentation methods suffer from limited segmentation accuracy due to high missed detection rates of small lesions and difficulties in balancing global and local features, an improved U-Net model integrated with attention mechanisms is proposed for brain white matter lesion segmentation. Using a 15-layer U-Net as the backbone, this model embeds a multi-scale feature calibration module into the encoder to simultaneously capture local details and global context via a dual-branch structure. Subsequently, a multi-stage feature enhancement module is designed to realize cross-layer feature fusion and enhance the perception capability of small lesions. Furthermore, a cross-layer feature integration module is constructed to build feature extraction pathways at multiple depths for refined multi-scale feature fusion. Finally, a boundary recall loss function is introduced to improve boundary segmentation precision by the Sobel gradient alignment mechanism. Experiments on the 2017 WMH segmentation challenge dataset and the brain white matter dataset from Tongji Hospital (Wuhan) show that the proposed method achieves Dice similarity coefficients of 0.831 8 and 0.854 5, and recall rates of 0.854 7 and 0.885 1, respectively. These findings confirm the effectiveness and precision of the proposed method in brain white matter lesion segmentation.