SFMANet: A Spatial-Frequency multi-scale attention network for stroke lesion segmentation.

Area of Science:

• Neuroimaging Analysis
• Medical Image Segmentation
• Artificial Intelligence in Medicine

Background:

• Accurate stroke lesion segmentation is vital for evaluating rehabilitation progress.
• Irregular shapes, unclear boundaries, and similar signal intensities complicate lesion differentiation from healthy tissue.

Purpose of the Study:

• To develop a novel deep learning method, SFMANet, for improved stroke lesion segmentation.
• To address challenges posed by irregular lesion shapes and blurred boundaries in neuroimaging data.

Main Methods:

• Proposed SFMANet, a UNet-based architecture incorporating Spatial-Frequency Gating Units (SFGU) and Dual-axis Multi-scale Attention Units (DMAU).
• SFGU enhances feature representation and utilizes redundant information; DMAU refines edge positioning using multi-scale context.
• An Information Enhancement Module (IEM) was integrated to minimize information loss and establish long-range dependencies.

Main Results:

• SFMANet demonstrated superior performance in capturing fine edge details of stroke lesions.
• Experiments on ISLES 2022 and ATLAS datasets showed SFMANet outperforming existing segmentation methods.

Conclusions:

• SFMANet offers an effective solution for accurate stroke lesion segmentation in neuroimaging.
• The proposed network architecture enhances the precision of lesion delineation, aiding in clinical assessment and rehabilitation planning.