STU3: Multi-organ CT Medical Image Segmentation Model Based on Transformer and UNet

Abstract

With the popularity of artificial intelligence applications in the medical field, U-shaped convolutional neural network (CNN) has garnered significant attention for their efficacy in medical image analysis tasks. However, the intrinsic limitations of convolutional operation, particularly in the receptive field since it is an end-to-end learning method, impede the establishment of long-term semantic feature dependence and holistic context information connection. This results in the edge contour details insensitive during the image segmentation task. To mitigate these shortcomings, Transformer architectures equipped with self-attention mechanism offer a potential alternative for encoding long-term semantic features and capturing global contextual information. Motivated by these insights, this paper proposes a novel U-shaped Transformer architecture, denoted as STU3, specifically engineered for medical image segmentation. Initially, a parallel training paradigm is employed that distinguishes between global fine-grained and local coarse-grained image features, optimizing the feature extraction process. Secondly, to alleviate the restrictions on fine-grained feature fusion due to peer skip connections, we propose a Residual Full-scale Feature Fusion module (RFFF) as the global decoder component. Lastly, a Global-Local Feature Fusion Block (GLFB) is implemented to seamlessly integrate the fine-grained and coarse-grained features, thereby constructing a comprehensive global information dependency network. This ensures a high level of accuracy in medical image segmentation tasks. Experimental evaluations conducted on abdominal and cervical multi-organ CT datasets substantiate the superiority of the proposed STU3 model over most current models, particularly in terms of the Dice Similarity Coefficient evaluation metric.