Abstract Body: Background: Clinical stratification for Abdominal Aortic Aneurysm (AAA) rupture relies on the maximum diameter criterion, an imperfect measure that fails to account for individual patient variability and leads to suboptimal surgical timing. While biomechanical metrics like Peak Wall Stress (PWS) and radiomic features offer superior predictive value, their clinical adoption is hindered by computationally expensive, time-consuming manual segmentation and Finite Element Analysis (FEA). We hypothesized that a single, multi-task 3D-CNN, trained on a public dataset, could unify these disparate approaches to create a rapid, automated, and superior "image-to-risk" pipeline.
Methods: A 3D-CNN model ("Aorto-FusionNet"), based on the nnU-Net architecture, was developed using the open-source AAA-100 dataset (N=100). For this training cohort, PWS "ground truth" maps were generated using FEA. The model was trained from a single CTA scan to perform three simultaneous tasks: (1) semantic segmentation of the lumen, intraluminal thrombus (ILT), and calcifications; (2) regression-based prediction of 3D PWS maps, bypassing FEA; and (3) extraction of a 512-dimensional latent radiomic feature vector. A final risk score classifier was trained and validated on a separate, retrospective external validation cohort (N=250) to predict 5-year adverse AAA events (rupture or repair).
Results: The Aorto-FusionNet segmentation head achieved mean Dice scores of 0.96 (Lumen), 0.91 (ILT), and 0.85 (Calcification). AI-predicted PWS strongly correlated with FEA-generated PWS (r=0.97, p<0.001), demonstrating its utility as an FEA surrogate. On the external validation set, the merged Aorto-FusionNet risk model achieved an Area Under the Receiver Operating Characteristic (AUROC) of 0.95 (95% CI: 0.92-0.98) for predicting 5-year events. This significantly outperformed the diameter-only model (AUROC 0.78, 95% CI: 0.73-0.82; p<0.001) and provided a Net Reclassification Improvement (NRI) of 0.42 (p<0.001). Feature importance analysis identified AI-PWS, diameter, and latent ILT textural features as the top three predictors.
Conclusion: A multi-task 3D-CNN can be successfully trained on public data to automate and unify segmentation, biomechanics, and radiomics from a single CTA scan. This "image-to-risk" model, by serving as a surrogate for time-consuming FEA, provides a comprehensive, automated risk score that improves upon the current diameter-based paradigm.