A Cartesian Grid Representation of Left Atrial Appendages for a Deep Learning Estimation of Thrombogenic Risk Predictors

Abstract

Computational fluid dynamics (CFD) simulations have recently been used to assess haemodynamic implications of atrial fibrillation on geometrical meshes built from patient-specific data. Some deep learning architectures, such as Fully Connected Networks (FCN), have demonstrated potential in accelerating CFD simulations, determining the relation between object geometry and model outcomes after finding correspondences with classical surface registration techniques. However, other successful architectures, such as Convolutional Neural Networks (CNN), have not been used yet in this application since geometrical meshes do not present a Euclidean structure, unlike medical images. The primary goal of this study was to estimate a fast surrogate of fluid simulations, based on a CNN architecture, for the prediction of thrombus formation risk in the left atrial appendage (LAA). For this purpose, a new flattened representation of the LAA was achieved by sampling its corresponding mesh in two directions: from the LAA junction to the left atrium (i.e. the ostium) to the tip, using the normalized gradient of the heat flow, producing radial isolines; and along the radial isoline direction, ordering the sampled points by their angle to a reference point. Using the resulting discretization, two FCN and one CNN architectures were tested. The CNN obtained the lowest mean absolute error and better predicted the areas of more elevated thrombogenic risk, while being orders of magnitude more computationally efficient than registration-based methods.