Left atrial appendage (LAA) clotting risk inferrence and flow reconstruction from 4D Contrast-CT imaging by Multi-Physics-Informed Neural Network (PINN)

Atrial fibrillation (AFib) is the most common cardiac arrhythmia affecting one in three people worldwide. The left atrial walls move weakly and irregularly during AFib, creating regions of blood stasis and thus thrombus may form inside the LAA. Patients with AFib have higher rates of atrial thrombosis and have five times increase in ischemic stroke risk incidence than healthy individuals. Current clinical risk scores for stroke in AFib patients are not patient-specific and have moderate accuracy. Moreover, 4D Contrast-CT images used to assess left atrial hemodynamics have motion artifacts during image reconstruction, which hinders the accuracy of secondary analysis derived from contrast dynamics.



We present LAA-PINN, a multi-physics-informed-neural-network approach that infers clotting risk inside the LAA and reconstructs the entire left atrial flow fields from partial or complete 4D Contrast-CT images. Its underlying physical models are Navier-Stokes, continuity, contrast transport equation, and residence time equation. We analyze LAA-PINN on CFD-generated ground-truth data and test the sensitivity of LAA-PINN vs. imaging parameters such as spatial and temporal resolution. Finally, we demonstrate the feasibility of using sinogram as training data to correct motion artifacts, infer clotting risk, and reconstruct flow fields all at once.