Automated CTA-Based Morphometric Analysis for Predicting Primary Entry Tear Location in Stanford Type A Aortic Dissection
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Preoperative localization of the primary entry tear in Stanford type A aortic dissection helps determine the extent of aortic repair but remains challenging on emergency computed tomography angiography (CTA). We developed and externally validated a model to distinguish ascending aortic from arch entry tears using automatically quantified aortic morphology.
Methods
This retrospective study included a single-center development cohort of 680 patients, comprising 529 with ascending and 151 with arch entry tears, and an independent ImageTAAD external cohort of 89 patients, comprising 47 with ascending and 42 with arch entry tears. Two three-dimensional nnU-Net models segmented the true lumen, patent false lumen, false lumen thrombus, and anatomical zones on preoperative arterial-phase CTA. We extracted 131 morphological features and used leakage-controlled five-fold cross-validation to select features and compare five classifiers. In the development cohort, intraoperative tear location served as the reference standard, supplemented by CTA review. The final model and deployment threshold were locked before external validation.
Results
Five features representing ascending false lumen and thrombus burden, arch geometry, and ascending aortic caliber were retained. Random forest achieved the highest internal cross-validated discrimination, with an area under the receiver operating characteristic curve (AUC) of 0.862 (95% CI, 0.833 to 0.892). At the pooled out-of-fold threshold of 0.281, sensitivity was 92.7% and specificity was 66.5%. In external validation, the locked model achieved an AUC of 0.785 (95% CI, 0.692 to 0.879). At the development-derived threshold of 0.253, sensitivity was 97.6% and specificity was 17.0%. In an exploratory post hoc analysis, a cohort-specific threshold yielded 78.6% sensitivity and 68.1% specificity.
Conclusions
A five-feature model automatically derived from preoperative CTA distinguished ascending from arch primary entry tears and retained discrimination externally. Site-specific threshold assessment and prospective multicenter validation are required before clinical use.
