Skip to main content

Resource > Aortas, Aneurysms, & AI

Aortas, Aneurysms, & AI

July 31, 2023

Aortas, Aneurysms, & AI

This animation shows the predicted labels for the hold-out dataset across 30 epochs of training alongside the plot of the performances of both sets.

Authors

Engineer Intern Sofia Lima, Principal John Mould, Ashley Peterson and Kristian Debus

Abstract

In this demonstration, the goal of the in silico trial is to assess the efficacy and safety of a generic, representative stent graft for the treatment of a thoracic aortic aneurysm (TAA). We used finite element analysis (FEA) for surgery simulation, along with Python packages for machine learning (ML) and data visualization. With such a valuable dataset, we aimed to reap the benefits of mesh-based representation in terms of efficiency and accuracy by using geometric deep learning (DL) with the raw surface meshes. We demonstrated a graph-level classification task where binary labels (surgical success or failure) are assigned based on endoleak presence. Our graph neural network (GNN) achieved similar accuracy on the validation set compared to our simpler ML models tested previously. In this report, we discussed the interpretability and performance of these approaches of varying complexity.