PCMA-UNet: a hybrid attention mechanism based on UNet3+ for stroke segmentation networks

Abstract Stroke is an acute cerebrovascular disease, which may lead to disability and death. The accurate segmentation of stroke lesions is of great significance for clinical diagnosis and patient prognosis. The existing deep learning mainstream segmentation network has made some progress in the direction of auxiliary diagnosis, but it lacks of attention to spatial information and location information, and cannot adapt well to lesions with different shapes and locations. Therefore, a hybrid attention mechanism of stroke segmentation network PCMA-UNet was proposed. With UNet3+ as the basic framework, pyramid Squeeze attention was used to extract spatial information of different scales in the feature map, and coordinate attention was introduced to strengthen the remote dependence relationship in the feature space and highlight the key information of lesion location. Multi scale attention was introduced at the end of the model to highlight the most significant feature maps in different scales to adapt to the adjustment of the current segmentation lesion size. Experiments were conducted on the cooperative hospital stroke dataset AIS and ISLES2022, respectively. The experimental results showed that the proposed method can effectively segment the stroke lesion region, improve the segmentation accuracy and provide a reliable basis for clinical diagnosis.