TransUNet＋: Redesigning the skip connection to enhance features in medical image segmentation

The new architecture TransUNet, which combines convolutional neural networks (CNNs) and transformers, has displayed competitive performance in medical image segmentation. In this paper, an efficient model named TransUNet＋, which can achieve promising results with a redesigned skip connection, is proposed for medical image segmentation. The redesigned skip connection contains an enhancement module, which can effectively enhance the skip features to improve global attention by using the score matrix of the transformer block. As the column vectors of the score matrix represent the relationship between the patches and the whole image, they can be used for feature enhancement. To validate the proposed TransUNet＋, series of experiments are performed based on three different medical image segmentation datasets covering multiple imaging modalities. Experimental results show that the proposed TransUNet＋ outperforms other state-of-the-art methods based on three datasets, and the proposed TransUNet＋ displays better performance in small organ segmentation.