LiU-Net: Ischemic Stroke Lesion Segmentation Based on Improved KiU-Net

Earlier and more accurate diagnosis of ischemic stroke is crucial in enhancing the therapeutic outcome for patients. CT technology currently stands as the most rapid diagnostic modality in clinical medicine. Due to the diverse and complex shape of ischemic stroke lesions, accurate segmentation remains a challenging task for automated diagnosis systems. This paper, proposes an ischemic stroke lesion segmentation network, LiU-Net. It based on KiU-Net, which improves network performance and is more suitable for practical lesion segmentation applications. Firstly, KiU-Net combines the undercomplete network U-Net and the overcomplete network Kite-Net. It can simultaneously learn both image detail features and global structural features. Secondly, LiU-Net combines the axial self-attention module with KiU-Net. The introduction of attention can make the network achieve both segmentation accuracy and efficiency. In addition, to improve the flexibility of axial self-attention, a gate factor is introduced within the module to encode information about spatial structure of image. Finally, to address the issue of gradient vanishing, we incorporated residual connection into the network to bolster the feature maps at each depth level and facilitate effective cross depth feature integration. Since there are few publicly available datasets of CT images of ischemic stroke in medical images. We applied to Longcheng District People's Hospital, and processed the obtained images to form a dataset of ischemic stroke. The experimental results shown, LiU-Net is more accurate in segmenting different shapes of ischemic stroke lesions. Compared with KiU-Net, LiU-Net improves the Dice, Acc, and mIoU metrics by 2.44%, 3.4%, and 3.89% respectively. Therefore, LiU-Net is highly suitable for ischemic stroke lesion segmentation, and effectively assist computers in this task.