TransVNet: A Network to Improve TransUNet Using VGG16

The method based on Convolutional Neural Network (CNN) can automatically and accurately segment objects in medical images. The combination of CNN and long-range dependency feature extraction techniques, such as TransUNet, can further improve segmentation performance. However, the segmentation performance of TransUNet is affected by the feature extraction ability of the CNN architecture. In order to improve the liver segmentation performance of TransUNet in abdominal CT images, this paper proposes a TransVNet architecture that combines CNN and Transformer using a symmetrical U-shaped architecture, and uses skip connections to connect features between corresponding convolution stages. To verify the effectiveness of TransVNet in segmenting the liver in abdominal CT, we compared its performance with other liver segmentation methods using 5-fold cross validation on the public dataset 3Dircadb. The experimental results show that the proposed architecture achieves the highest average DICE score of 95.29%.