Patient-specific modeling of hemodynamic disorders using Physics Informed Neural Networks

Hemodynamic disorders are diseases caused by the altered dynamics of the blood flow in the circulatory system, due to anomalies like aneurysms or plaque deposition, and are one of the leading causes of preventable deaths in the US. Since these disorders are highly patient-specific in nature, building individualized hemodynamic profile models has gained a lot of interest. However, running large ensembles of these realizations with full 3D CFD (Computational Fluid Dynamics) models is computationally expensive. While traditional Data-driven techniques (such as Deep Learning) are computationally cheaper for inference, they typically require large amounts of high-fidelity data for training. Physics Informed Neural Networks (PINNs) overcome this challenge by providing a framework to leverage the underlying knowledge of the governing equations into training the neural network, thereby reducing the need for copious amounts of labelled training data to devise a biologically/physiologically consistent model. In this work, we demonstrate the feasibility of using PINNs to study hemodynamic flow conditions for arteriosclerosis and aneurysms, enabling faster, reliable evaluations of flow patterns for surgical planning and exploring possible medical intervention strategies (such as stents).