Superimposed Attention Enables TransUNet for Medical Image Segmentation

Accurately identifying and localizing lesion regions and extracting key structures and organs from medical images are essential for diagnosis, treatment planning and tracking disease progress. Existing CNN - Transformer models are still insufficient for extracting detailed features. To address this problem, a deep medical image segmentation model framework, SA-TransUNet, is proposed. Using TransUNet as the core model, deep separable convolution (DSC) is applied to enhance the learning ability of detailed features after CNN-Transformer extraction. In the decoding stage, a module called Deep Convolutional Attention (DCCS) is introduced after feature fusion in the jump connection, which combines deep convolution from the perspective of expanding the sensory field to enhance the learning ability, and combines two attentional mechanisms (Convolutional Block Attention Module and Scaling Dot Product Attention) from the perspective of channel-space. Experimental results on the Synapse multi-organ segmentation dataset show that the SA-Trans UN et network outperforms other networks.