Image Segmentation Algorithm of Ischemic Stroke based on Wavelet Pooling Neural Network

Ischemic stroke is a high-risk brain disease that causes disability or death in adults worldwide. Rapid diagnosis and treatment, as well as segmentation of ischemic lesions from stroke medical images so that doctors can evaluate and develop treatment plans, are crucial to reducing brain damage. Currently, doctors' manual segmentation of lesions is time-consuming and labor-intensive and prone to missed or misdiagnosed diagnoses, delaying the best time for treatment. A well-trained neural network is considered as an effective aid for image segmentation tasks. However, in the traditional neural network down-sampling process, the pooling layer loses feature information, and the lost information significantly impacts the edge regions. While for medical images, edge regions are crucial features in the segmentation process. Therefore, to address the above issues, in this paper we improve upon the existing stroke lesion segmentation algorithm U-net, we design a segmentation network based on wavelet pooling and attention mechanisms. The network was tested on ISLES 2018 and ISLES 2022 datasets, and the Dice coefficient improved by 2.13% and 1.93% compared to baseline network, respectively.