UTAC-Net: A Semantic Segmentation Model for Computer-Aided Diagnosis for Ischemic Region Based on Nuclear Medicine Cerebral Perfusion Imaging

Abstract

Cerebral ischemia has a high morbidity and disability rate. Clinical diagnosis is mainly made by radiologists manually reviewing cerebral perfusion images to determine whether cerebral ischemia is present. The number of patients with cerebral ischemia has risen dramatically in recent years, which has brought a huge workload for radiologists. In order to improve the efficiency of diagnosis, we develop a neural network for segmenting cerebral ischemia regions in perfusion images. Combining deep learning with medical imaging technology, we propose a segmentation network, UTAC-Net, based on U-Net and Transformer, which includes a contour-aware module and an attention branching fusion module, to achieve accurate segmentation of cerebral ischemic regions and correct identification of ischemic locations. Cerebral ischemia datasets are scarce, so we built a relevant dataset. The results on the self-built dataset show that UTAC-Net is superior to other networks, with the mDice of UTAC-Net increasing by 9.16% and mIoU increasing by 14.06% compared with U-Net. The output results meet the needs of aided diagnosis as judged by radiologists. Experiments have demonstrated that our algorithm has higher segmentation accuracy than other algorithms and better assists radiologists in the initial diagnosis, thereby reducing radiologists’ workload and improving diagnostic efficiency.