On the Use of Machine Learning to Predict Rupture Potential Index for Abdominal Aortic Aneurysms
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The overall mortality rate due to rupture of an abdominal aortic aneurysm (AAA) is greater than 80%. The current clinical standard for assessing rupture risk of an AAA is a size-based approach solely based on measuring the maximum transverse diameter (Dmax) of the aneurysm. If Dmax is greater or equal than 5.0 cm (in the U.S.), repair is recommended; otherwise, the patient is placed in a surveillance program consisting of "watchful waiting" with periodic imaging follow ups. This approach is not always reliable as some small aneurysms rupture prior to reaching the aforementioned critical diameter. Likewise, some large AAAs remain stable and are not diagnosed until they have exceeded the critical size. Biomechanical measures such as peak wall stress and 99th percentile wall stress have been shown to be better predictors of AAA rupture compared to Dmax. AAA rupture is a mechanical event that occurs when the stress on the AAA wall exceeds the wall strength. Therefore, the rupture potential index (RPI) – the ratio of local wall stress to wall strength ‒, is a biomechanical metric that can be used as an AAA rupture risk predictor. The calculation of RPI is based on patient-specific finite element analysis (FEA), which requires prior knowledge of volume meshing, use of a FEA solver, and warrants patient-specific material properties. Moreover, FEA can be user-dependent and the outcomes vary with mesh density and the applied boundary conditions. Geometric markers such as area-averaged Gaussian curvature and minimum wall thickness, have been previously utilized as surrogates to predict spatially averaged wall stress (SAWS) in symptomatic and emergently repaired AAAs. Using these geometric markers, AAA wall stress can be predicted in lieu of FEA. Similarly, area-averaged Mean curvature and proximal neck diameter have been quantified to accurately predict SAWS in asymptomatic and electively repaired AAAs. In this study, we built a machine learning model for the prediction of RPI using geometric markers derived from 3D reconstructed computed tomography angiography images. The current protocol for AAA geometric quantification calculates a set of global markers representative of the shape, size, wall thickness, and curvature of the AAA sac. We quantified the spatial distribution of such geometric markers for a more comprehensive prediction of RPI. Further, the predicted RPI was utilized to identify and classify high-risk AAA. In summary, we developed and validated an algorithm that can predict RPI in AAA models using a machine learning approach to identify geometric markers from individual clinical images.