IMPROVED TRANSUNET WITH COORDINATE ATTENTION FOR TRUE AND FALSE LUMEN SEGMENTATION IN TYPE-A AORTIC DISSECTION

Type A aortic dissection is a serious cardiovascular disease that requires early diagnosis and timely intervention. Accurate segmentation of true cavities and false cavities in computed tomography angiography (CTA) images is crucial for clinical decision-making. In this study, we proposed and implemented an improved TransUNet-based segmentation model. Inspired by the nondimensionality reduction design of ECANet channel attention, this model proposes an improved coordinate attention mechanism to enhance the feature distinction between lumens. Furthermore, we propose an Edge-Sensitive Dice (ES Dice) loss, which integrates Dice loss, binary cross-entropy, and edge-aware terms to better capture boundary details. Our model was trained and evaluated on 9831 2D CTA slices extracted from the 3D CT volumes of 40 TAAD patients, with each slice manually annotated for TL and FL regions. Experimental results show that our method achieved a Dice coefficient of 0.9004, significantly outperforming several state-of-the-art baseline methods. In future work, we plan to extend the method to 3D volumetric data and perform multi-center validation to support broader clinical applications.