Abstract
In neuroimaging analysis, accurately delineating stroke lesion areas is crucial for assessing rehabilitation outcomes. However, the lesion areas typically exhibit irregular shapes and unclear boundaries, and the signal intensity of the lesion may closely resemble that of the surrounding healthy brain tissue. This makes it difficult to distinguish lesions from normal tissues, thereby increasing the complexity of the lesion segmentation task. To address these challenges, we propose a novel method called the Spatial-Frequency Multi-Scale Attention Network (SFMANet). Based on the UNet architecture, SFMANet incorporates Spatial-Frequency Gating Units (SFGU) and Dual-axis Multi-scale Attention Units (DMAU) to tackle the segmentation difficulties posed by irregular lesion shapes and blurred boundaries. SFGU enhances feature representation through gating mechanisms and effectively uses redundant information, while DMAU improves the positioning accuracy of image edges by integrating multi-scale context information and better allocates the weights of global and local information to strengthen the interaction between features. Additionally, we introduce an Information Enhancement Module (IEM) to reduce information loss during deep network propagation and establish long-range dependencies. We performed extensive experiments on the ISLES 2022 and ATLAS datasets and compared our model's performance with that of existing methods. The experimental results demonstrate that SFMANet effectively captures the edge details of stroke lesions and outperforms other methods in lesion segmentation tasks.