Medical Image Segmentation Based on Transformer and Spatial Recursive Convolution

Artificial intelligence-assisted automated medical image segmentation plays a vital role in helping clinicians make accurate diagnoses and develop effective treatment plans. However, as a significant machine learning model in medical image processing, a U-shaped network (U-Net) faces challenges in capturing long-range dependencies, particularly in medical images with complex textures where structures often blend into the background. To address these challenges, we propose the Transformer and Spatial Recursive U-Net (TSU-Net), a novel architecture that integrates the transformer technology and the spatial recursive convolution, building upon the UNet framework. The main contribution of TSU-Net lies in the adaptive transformer (AT) block, which integrates transformer mechanisms with adaptive pooling to effectively capture longrange dependencies and enhance multi-level abstract features. Furthermore, we introduce the spatial recursive convolution (SRC) block, which iteratively updates features across layers, thereby improving the network’s capacity to model spatial correlations and describe intricate features in medical images. Experimental results on a cardiac segmentation dataset demonstrate that TSU-Net can enhance segmentation accuracy, underscoring its potential for medical image segmentation tasks.