CLAC-Net: a composite medical image segmentation framework using self-attention and cross-layer asymmetric connections

Medical image semantic segmentation plays a crucial role in the localization of organs and lesions, analysis and quantification of pathologies, and surgical planning and navigation. Existing network models often encounter several challenges: (1) Difficulty in matching and integrating context semantic information effectively. (2) Limited ability to capture global information and combine it with contextual data. (3) Semantic disparities caused by the use of upsampling and downsampling in the model. To address these issues, we proposed the Cross-layer and Asymmetric-connection Net (CLAC-Net), an adaptation of the custom-built baseline structure ACUNet. The design considerations for CLAC-Net were as follows: a. Asymmetric skip connections were designed to reduce feature loss caused by large sequence spans, enhancing the match of semantic information between preceding and following contexts. b. The Cross-Layer Relay Knot System (CLRKS) was utilized to capture global information and integrate it with contextual data. c. The Deep Embedding Attention Bottleneck (DEAB) was developed to strengthen semantic transmission between the encoder and decoder, thereby reducing the semantic gap. Experimental results have demonstrated the superiority of our model in comparative analyses. Our Code is available at: https://github.com/YF-W/CLAC-Net.