UCATR: Based on CNN and Transformer Encoding and Cross-Attention Decoding for Lesion Segmentation of Acute Ischemic Stroke in Non-contrast Computed Tomography Images.

The acute ischemic stroke (AIS) impacts extensively all over the world, the early diagnosis can provide valuable property information of disease. However, it's difficult for our human eyes to distinguish the fine pathological changes. Here we introduce self-attention mechanisms and propose UCATR, an NCCT image segmentation network for AIS lesions. It uses the advantages of Transformer to effectively learn the global context features of the image, and is based on convolutional neural network (CNN) and Transformer as the encoder, adding Multi-Head Cross-Attention (MHCA) modules to the decoder to achieve high-precision spatial information recovery. This method is experimentally verified on the NCCT dataset of AIS provided by Chengdu Medical College in China to obtain that the Dice similarity coefficient of lesion segmentation is 73.58%, which is better than U-Net, Attention U-Net and TransUNet. Furthermore, we conduct ablation study on the MHCA module at three different positions in the decoder to prove its efficiency.