Collaborative transformer U-shaped network for medical image segmentation

Abstract

Recent advances in the Transformer have shown significant ability to understand the relationship between the lesion area and surrounding tissue, especially for medical image analysis. Existing medical image segmentation algorithms based on transformers often suffer from limited feature extraction granularity and overlook the semantic relationships between multi-scale features. To solve the above limitations, we propose CoTransUNet: a collaborative transformer U-shaped network, that effectively captures fine-grained features and long-range dependencies by performing context extraction between multiple scales. The designed Correlation Extraction (CE) module bridges the encoder and decoder to achieve effective interaction and information transfer. Specifically, a collaborative mechanism in the encoder is proposed that can efficiently exploit inductive bias to extract local fine-grained features of the image while having the ability to capture long-distance feature dependencies. Besides, the CE module focuses on deeply integrating contextual information of multi-scale features, which enriches feature representation by exploiting the intrinsic correlation between features at different scales. It can extract not only local and global features but also capture semantic information related to different multi-scale features simultaneously. Compared to TransUNet, CoTransUNet achieves a 4.91% improvement in DSC on the Synapse multi-organ segmentation dataset while using only a quarter of the parameters. The extensive experiments on three datasets, including skin lesion segmentation (ISIC2016, ISIC2017, ISIC2018) demonstrates that CoTransUNet achieves DSC scores of 92.18%, 85.59%, and 88.75%, respectively, and on Synapse multi organ segmentation achieves DSC score of 82.39% , which outperforms the baseline and other promising methods.