Anatomically Guided Two-Stage 3D Aorta Segmentation in CT Angiography

Abstract

Accurate segmentation of the aorta and its 23 substructures is crucial for planning endovascular interventions and treating aortic pathologies. However, deep learning approaches often face practical limits when handling high-resolution 3D computed tomography angiography (CTA). We propose a two-stage framework, AortaSeg, which integrates a 2D object detector with a 3D patch-based segmentation network to reduce memory usage without sacrificing accuracy. Stage 1 localizes anatomical regions using a YOLOv7-based detector guided by reference organs from TotalSegmentator, and Stage 2 applies section-specific 3D DynUNets to cropped sub-volumes. This modular design enables efficient processing of large CTA volumes and improves focus on challenging substructures. On the hidden test set of the MICCAI 2024 AortaSeg Challenge, AortaSeg achieved a Dice Similarity Coefficient (DSC) of 0.767 +/- 0.032 and a Normalized Surface Distance (NSD) of 0.797 +/- 0.037, ranking 5th among 32 teams, demonstrating a favorable balance between performance and efficiency.