Reducing the Cost of MRI Scanning by Enhancing Data with Underlying Physics via Physics-Informed Neural Networks
DescriptionPhysics informed neural networks (PINNs) have grown from a toy machine learning concept to very powerful and applicable in a variety of modern problems. As PINNs continue to grow, they will inevitably be utilized more in the fields of medicine and biology, where not only are the domains of interest incredibly complex but information on them incomplete. We study PINN performance in said area via the lens of aortic dissections (AD) informed by MRI scanning. Specifically, we consider the prediction accuracy of PINNs as a function of 4D flow MRI in both spatial and temporal resolutions, and consider PINN prediction of gradient based parameters, wall shear stress and shear rate. Three AD aneurysms are analyzed, ones with large, medium, and small sized mouths. These mouths lack any boundary conditions making this an unsolvable problem with standard computational techniques. We utilize PINN aided by 2D MRI data to learn the hemodynamics of the domain. We conclude that full MRI resolution may not be required, saving on scanning cost, and in the case of AD aneurysms, larger mouths lead to more accurate results due to the larger order of magnitude in the velocities which are easier for PINNs to learn.
TimeTuesday, June 2712:00 - 12:30 CEST