Estimating coronary hemodynamics with a Deep Learning-based approach

Atherosclerosis is the primary pathology impacting the coronary arteries, generating stenosis which constrict blood flow. It is a chronic disease characterized by the thickening of the coronary wall from fat buildup. In coronary arteries atherosclerosis leads to coronary artery disease (CAD) with myocardial infarction as deadliest complication. CAD is responsible for 30% of fatalities in individuals aged 35 and older. Evaluating the severity of coronary lesions is crucial in identifying the appropriate treatment. Therefore, the physiological assessment of CAD has gained increasing importance in clinical and research settings. In clinical practice, the traditional technique for CAD diagnosis, coronary angiography imaging, has been complemented by flow-based and pressure-based functional quantities. However, such quantities only moderately predict the risk associated with the presence of non-obstructive lesions. Therefore, interest has grown on the role of local hemodynamic disturbances in the onset and progression of coronary artery lesions. The mechanical action exerted by fluid forces on the endothelium, quantifiable in terms of wall shear stress (WSS) and WSS-related measures, has been identified as involved in coronary atherosclerotic plaques progression and rupture. WSS is typically computed by means of computational fluid dynamics (CFD) simulations, a valuable resource for patient-specific, non-invasive assessment of hemodynamics in arteries. However, CFD simulations are associated with high computational time. Therefore, the use of Deep Learning (DL) techniques has been explored to accelerate CFD simulations. To this end, the aim of this thesis is to investigate the application of DL-based methods for the swift, “real time” estimation of WSS on patient-specific 3D coronary artery geometries obtained from quantitative coronary angiography. In my Master Thesis project a dataset of 187 patient-specific stenosed coronary artery models extracted from coronary angiography was considered. Gauge Equivariant Mesh Graph Convolutional Network (GEM-GCN) was adopted as geometric DL model to predict WSS, with CFD results serving as ground truth. GEM-GCN is based on Graph Neural Networks (GNNs), acting on surface meshes and exploiting anisotropic kernels and input features to correlate the model shape and the quantity to predict. Firstly, the dataset of patient-specific coronary arteries was used for an external validation of the GEM-GCN previously trained with a dataset of 1600 synthetic idealized models. Then, the same dataset was used to train and test the DL model, tuning the hyperparameters and subdividing the dataset into training (80%), validation (10%) and test set (10%). The external validation step highlighted the difficulty of the proposed DL model to generalize to realistic data. However, the predictions of the network trained with the dataset of patient-specific coronary arteries showed that the DL model captures the main hemodynamic features in terms of WSS magnitude and direction, but not for all coronary models. The average percentual error in WSS magnitude on the single model ranged from 11.90% to 56.41%, while the average error in WSS direction ranged from 2.69° to 3.87°. In conclusion, this study displays the potential of DL algorithms as CFD surrogate models for real time hemodynamics parameters estimation. While generalizable results are yet to be achieved, future models will benefit from improvements in dataset size and representativeness of geometrical complexity.