AI-Based Segmental Volumetry of the Downstream Aorta in Aortic Dissection: End-to-End Versus Hybrid Strategies

Automated segmental volumetry is effective for aortic dissection surveillance but lacks standardization. This study aimed to identify the optimal deep learning workflow by comparing a direct "end-to-end" approach with a "hybrid" strategy integrating vertebral-based partitioning. We analyzed 112 computed tomography scans (57 type A, 55 type B). The aorta was subdivided into five segments using vertebral landmarks. We evaluated two strategies across three architectures (nnU-Net, SwinUNETR, and U-Mamba): (1) end-to-end (simultaneous multi-class segmentation) and (2) hybrid (true and false-lumen segmentation followed by vertebral-based partitioning). Performance was assessed using the dice similarity coefficient, 95% Hausdorff distance, and relative volume error. The hybrid strategy improved segmental accuracy over the end-to-end approach across architectures. Using nnU-Net, segment-averaged true lumen dice similarity coefficient improved from 0.877 to 0.945 (p < 0.001), and 95% Hausdorff distance decreased from 16.87 mm to 4.41 mm. False-lumen accuracy also improved but was lower than the true lumen and declined in the distal segments. The hybrid strategy also reduced false-positive false-lumen detections in negative cases (nnU-Net, from 50.0% to 29.2%). Bland‒Altman analysis showed low bias but wide limits of agreement (approximately ± 40%), narrower for the hybrid strategy in the false lumen. For automated segmental volumetry in aortic dissection, a hybrid workflow-combining lumen segmentation with vertebral-based partitioning-improved segmental boundary accuracy over an end-to-end approach and provided a standardized, reproducible partitioning across same-patient scans. Further improvement of the underlying segmentation will be needed before routine clinical surveillance, toward which this workflow offers a reproducible foundation.