Abstract 16111: Clinical Application of Deep Learning for Stress Analysis of Ascending Thoracic Aortic Aneurysms

Abstract

Introduction: Current guidelines for elective surgery of ascending thoracic aortic aneurysms (aTAAs) use aneurysm size as primary determinant for risk stratification of adverse events. Biomechanically, dissection may occur when wall stress exceeds wall strength. A widespread method for stress analysis is structural finite-element analysis (FEA). Patient-specific aortic geometries are easily obtainable and stress distributions can potentially predict risk of dissection. However, FEA is a time-consuming and difficult procedure. To bypass this issue, a recent study has developed the first deep learning (DL) approach for a fast and accurate estimation of aortic wall stress distributions. Hypothesis: In this study, we assessed the hypothesis that this deep learning approach can be applied to a large clinical dataset. Model performance was measured by comparing FEA and DL stress predictions in parallel. Methods: Patients with aTAA (n = 169) were studied. Patient-specific aneurysm geometries were obtained from ECG-gated computed tomography. Shapes were represented by hexahedral meshes with 9648 nodes and 6336 solid elements. FEA peak wall stresses and stress distributions were determined using LS-DYNA software with user-defined fiber-embedded material models under systolic pressure. The DL model was implemented in Julia and consisted of unsupervised and supervised learning algorithms. Training was performed on a training set of 152 shapes and testing set of 17 shapes with 10-fold cross-validation. Mean absolute error (MAE) and absolute error of peak stress values (APE) were used to compare DL model predictions with FEA values considered to be ground truth. Results: Average stress values predicted by our DL model were 175.64 ± 4.17 kPa and 95.69 ± 2.15 kPa in the circumferential and longitudinal direction, respectively. We computed a MAE of 5.06 ± 1.08 kPa and APE of 2.58 ± 1.39 kPa in the circumferential direction and MAE of 4.51 ± 0.98 kPa and APE of 2.32 ± 1.84 kPa in the longitudinal direction. Conclusions: DL model trained exclusively on clinical data was able to accurately predict stress distributions on complex aortic geometries. Fast and accurate stress predictions will facilitate real-time clinical applications for the risk assessment of aTAAs.