Liver Image Segmentation Method Based on TransUNet and Axial Attention

In the task of liver image segmentation, due to the complex edges and internal structure of liver images, as well as the presence of adjacent organs, conventional medical image segmentation networks have exhibited local deficiencies in handling long-range dependencies and multi-scale information. These deficiencies may impact the accuracy of liver segmentation models. To address the small receptive fields of convolutional operations in previous medical image segmentation models and the information loss issue in the skip connections of the TransUNet model, this paper introduces Axial Attention and residual connections into the skip connections of the TransUNet model. Axial Attention weights the importance of different dimensions, better capturing and enhancing liver edges and other targets. Additionally, the use of residual connections effectively alleviates gradient vanishing and information loss issues. Experimental results demonstrate that the improved TransUNet model proposed in this paper achieves dice and IoU evaluation metrics of 97.2% and 94.0%, respectively, on the LiTS17 dataset, representing an improvement of 1.6 percentage points and 2.1 percentage points in Dice and IoU metrics, respectively, compared to the original TransUNet model. Its liver segmentation performance surpasses that of the original TransUNet model, and the improved TransUNet model can more accurately identify liver edges, enhancing feature representation and information flow, and providing more refined segmentation results.