SDLU-Net: A similarity-based dynamic linking network for the automated segmentation of abdominal aorta aneurysms and branching vessels

Abstract

Accurate preoperative measurement of abdominal aortic aneurysms (AAAs) and associated vascular structures is indispensable for surgical planning. However, the precise extraction of complex multi-branch vascular structures from CTA images presents significant challenges. These challenges stem from the high individual variation in vascular structure, the difficulty of separating important organ branches from the surrounding tissue area, and the unclear boundaries between main branches and organ branches. To overcome these obstacles, the SDLU-Net network architecture is introduced. Firstly, a novel spatial positional encoding structure, named “dual position encoding”, is presented to preserve the spatial relationships of vascular anatomical structures throughout the network operations. Secondly, a similarity-based dynamic attention linking module is proposed to effectively distinguish vessels from tissues, and the main branches from other branches. Furthermore, dual skip connections and channel attention modules are integrated into the decoding layer to enhance the information flow within the network. Lastly, a hybrid loss function was introduced to address the issue of sample imbalance, thereby improving the segmentation results of organ branches. Extensive experiments have demonstrated that SDLU-Net shows excellent segmentation performance in multi-branch blood vessels of the abdominal aorta, highlighting its substantial clinical potential for surgical planning involving complex multi-branch structures.