Quantifying three-dimensional aortic morphology is important for surgical planning in thoracic endovascular aortic repair (TEVAR). There is no agreed standard for how best to represent aortic shape for predicting outcomes. Two broad paradigms are commonly used: data-driven statistical shape analysis (SSA), which uses statistical methods and dimensionality reduction to capture dominant shape modes, and geometrically-informed descriptors grounded in differential geometry, such as curvature measures. This study directly compares these paradigms on a clinical imaging cohort to assess interpretability, robustness, and generalizability.
The analysis used a cohort of 290 computed tomography angiography (CTA) scans categorized by surgical outcome into non-pathological, successful TEVAR, and failed TEVAR. The authors report that patient imaging data were obtained under University of Chicago Institutional Review Board approvals IRB20-0653 and IRB21-0299, and that participant consent procedures and de-identification complied with institutional guidelines and the Declaration of Helsinki. The manuscript is a preprint and has not undergone peer review; the authors caution it should not be used to guide clinical practice.
The geometrically-informed approach reduced aortic shape to a two-dimensional feature space comprised of mean aortic radius (R) and the normalized fluctuation in integrated Gaussian curvature (˜δK). These descriptors are grounded in differential geometry: R captures average global size, while integrated Gaussian curvature quantifies intrinsic surface bending behavior. The authors emphasize that these descriptors are directly interpretable in anatomical and geometric terms, offering potential clinical transparency when used for classification.
For the SSA paradigm the authors constructed a point-cloud shape model of the aorta from the segmentation masks and applied Principal Component Analysis (PCA) for dimensionality reduction. Classification performance for SSA was evaluated as a function of the number of retained principal components (PCs). The leading PCs were found to encode global size variation among shapes.
Classification experiments used Gaussian process classification with 10-fold cross-validation on the training cohort. The geometrically-informed two-feature model (R, ˜δK) was compared against SSA models retaining varying numbers of principal components, including evaluations up to 20 PCs. Generalizability was further tested on an independent validation dataset (GORE TAG 08-01), which was accessible to the authors under a data transfer agreement and is not publicly available.
The study found that SSA’s leading principal components reflected global size differences and were statistically redundant with the pair (R, ˜δK). Despite encoding similar information, the PCs did not have a one-to-one mapping to interpretable anatomical quantities, limiting their mechanistic transparency.
In cross-validated classification on the training cohort, the geometrically-informed feature pair achieved a higher weighted F1 score than SSA when SSA was constrained to retain up to 20 principal components. Although SSA achieved greater than 90% accuracy on the full dataset as dimensionality increased, the authors attribute this apparent performance gain to overfitting rather than to a genuine increase in discriminative power. Evaluation on the independent validation dataset showed that the curvature-and-radius descriptors generalized better than the PCA-based SSA models.
These findings support the authors’ central claim that geometrically-informed descriptors offer improved interpretability and robustness relative to higher-dimensional, data-driven shape representations in this context.
The authors state that de-identified training cohort data, including segmentation masks and the derived shape descriptors used in the study, are publicly available at the project GitHub repository: https://github.com/SurgBioMech/plos_data. The independent GORE TAG 08-01 validation dataset is not publicly available; access was provided to the University of Chicago under a data transfer agreement and remains subject to that agreement’s terms. The authors declared no competing interests.
Within the reported cohort and analytic framework, curvature-based geometric descriptors combined with mean radius provided a more interpretable and better-generalizing framework for classifying aortic disease outcomes than PCA-based statistical shape analysis. The authors conclude that geometrically-informed descriptors are more robust and clinically translatable for aortic disease classification, while cautioning that SSA models with large numbers of components may overfit and lack direct anatomical interpretability. Because this work is a preprint, the authors note it has not been peer reviewed and should not be used to change clinical practice without further validation.