A Modular Deep-Learning Pipeline for Automated Aorta Characterization on CT

Accurate assessment of the aorta is critical for diagnosing and managing life-threatening conditions such as aneurysms, dissections, and connective tissue disorders. Manual measurements are prone to variability and are time-consuming. We propose a modular three-stage deep-learning pipeline for the automated characterization of the thoracic aorta and aortic root on computed tomography. The first stage detects regions of interest. The second stage uses a semantic segmentation module to isolate the thoracic aorta and extract the maximal diameter. The third stage employs a multi-task network to segment the aortic root and localize key landmarks, enabling precise measurement of the maximal aortic root diameter. Our method achieves mean Dice scores of 0.94 (thoracic aorta) and 0.96 (aortic root), a mean landmark localization error of 1.69 mm, and total inference times under 50 s on consumer-grade hardware. In a study involving three expert observers (30 cases), Bland–Altman and intra-class correlation analyses demonstrate that our tool yields measurements comparable to expert annotations (ICC ). The pipeline’s low computational footprint, anatomical focus on the aortic root, strong reproducibility, and high accuracy make it well suited for integration into diagnostic and pre-interventional workflows.