A cartesian grid representation of left atrial appendages for deep learning estimation of thrombogenic risk predictors

Abstract

Atrial fibrillation is the most common sustained arrhythmia in the world. This condition is responsible for about 20% of the cardioembolic ischemic strokes, of which, most are caused by thrombus formed in the Left Atrial Appendage (LAA). Computational fluid simulations have been used to assess hemodynamic implications of AF on a patientspecific basis. Deep Learning (DL) has shown potential in accelerating these simulations, although many of the most successful algorithms such as Convolutional Neural Networks (CNN) rely on the Euclidean structure of the data. Therefore, the aim of this study consisted on generating a fast surrogate of fluid simulations, that predicted the thrombus formation risk. For this purpose, a new flattened representation of the LAA was achieved by sampling the LAA in two directions: from the junction to the left atrium (i.e. ostium) to the tip (i.e. apex), using the normalized gradient of the heat flow and radially. It showed a great potential for clinical use and representation of LAA structures as a twodimensional flat geometry as well. Using the node discretization provided by the flattening algorithm, two Deep-Neural Networks (DNN) and one CNN configurations were tested. The mean absolute errors on the thrombogenic risk index given by the DL configurations were similar (0.74 and 0.73) while it was lower for the CNN (0.63). Similarly, the CNN-based approach detected better the highly thrombogenic risk areas compared to the DNN-based ones (87.9% vs 81.0% and 81.7%, respectively).