Non-Invasive Cardiac Hemodynamics : Reconstructing Left Atrial Appendage Blood Flow from CT Images using Physics-Informed Neural Operators

Sammanfattning

Computational fluid dynamics (CFD) simulations have many applications but are typically expensive and time intensive. One application is modeling the full blood-flow velocity field inside the heart from patient-specific geometries extracted from computed tomography (CT) imaging over time. This is a potentially useful non-invasive modality for clinicians selecting treatment for patients with atrial fibrillation, who are at an increased risk of blood-clot formation due to stasis in the left atrial appendage (LAA). This thesis explores whether neural operators trained on CFD data can act as surrogate models that replace CFD inference for predicting LAA blood flow more quickly while maintaining prediction accuracy. Several autoregressive Geometry-Informed Neural Operator (GINO) models were trained on CFD simulations from 29 patient geometries using different combinations of data and Navier–Stokes physics losses, and tested on 7 unseen geometries. The best performing base model was additionally fine-tuned at inference time on each test patient using a data-free, physics-only loss to measure potential accuracy improvements. No model achieved accuracy at a level where it could replace the underlying CFD simulations. However, including the continuity residual in the loss improved rollout stability and reduced both physics residuals, while adding the momentum residual was found to degrade performance. Inference-time fine-tuning consistently reduced both the error against the ground-truth flow field and the physics residuals across all test patients.