Transfer learning on physics-informed neural networks for tracking the hemodynamics in the evolving false lumen of dissected aorta

Abstract

Aortic dissection is a life-threatening event that is responsible for significant morbidity and mortality in individuals ranging in age from children to older adults. A better understanding of the complex hemodynamic environment inside the aorta enables clinicians to assess patient-specific risk of complications and administer timely interventions. In this study, we propose to develop and validate a new computational framework, warm-start physics-informed neural networks (WS-PINNs), to address the limitations of the current approaches in analyzing the hemodynamics inside the false lumen (FL) of type B aortic dissection vessels reconstructed from apolipoprotein null mice infused with AngII, thereby significantly reducing the amount of required measurement data and eliminating the dependency of predictions on the accuracy and availability of the inflow/outflow boundary conditions. Specifically, we demonstrate that the WS-PINN models allow us to focus on assessing the 3D flow field inside FL without modeling the true lumen and various branched vessels. Furthermore, we investigate the impact of the spatial and temporal resolutions of MRI data on the prediction accuracy of the PINN model, which can guide the data acquisition to reduce time and financial costs. Finally, we consider the use of transfer learning to provide faster results when looking at similar but new geometries. Our results indicate that the proposed framework can enhance the capacity of hemodynamic analysis in vessels with aortic dissections, with the promise of eventually leading to improved prognostic ability and understanding of the development of aneurysms.