Segmentation of acute ischemic stroke lesions based on deep feature fusion

Acute ischemic stroke (AIS) is a common brain disease worldwide, and diagnosing AIS requires effectively utilizing information from multiple Computed Tomography Perfusion (CTP) maps. As far as we know, most methods independently process each CTP map or fail to fully utilize medical prior information when integrating the information from CTP maps. Considering the characteristics of AIS lesions, we propose a method for efficient information fusion of CTP maps to achieve accurate segmentation results. We propose Window Multi-Head Cross-Attention Net (WMHCA-Net), which employs a multi-path U-shaped architecture for encoding and decoding. After encoding, multiple independent windowed cross-attentions are used to deeply integrate information from different maps. During the decoding phase, a Channel Cross-Attention (CCA) module is utilized to enhance information recovery during upsampling. We also added a segmentation optimization module to optimize low-resolution segmentation results, improving the overall performance. Finally, experimental results demonstrate that our proposed method exhibits strong balance and excels across multiple metrics. It can provide more accurate AIS lesion segmentation results to assist doctors in evaluating patient conditions. Our code are available at https://github.com/MTVLab/WMHCA-Net.