Predicting 4D Mitral Regurgitation Hemodynamics from Sparse Data using Hybrid Deep Operator Networks

Accurate quantification of mitral regurgitation (MR) is essential for clinical decisions and requires analyzing complex hemodynamics. 2D Particle Image Velocimetry (PIV) produces spatially sparse data, and transient Computational Fluid Dynamics (CFD) simulations remain computationally prohibitive for routine clinical use. This work proposes a hybrid Deep Operator Network (DeepONet) framework to reconstruct full transient volumetric (4D) hemodynamics from sparse 2D planar velocity slices and time-resolved pressure measurements. The model learns a generalized solution operator from Unsteady Reynolds-averaged Navier-Stokes (URANS) simulations and uses rapid Test-Time Adaptation (TTA) to fine-tune predictions directly on case-specific sparse observations. This approach infers plausible 3D jet trajectories in minutes. Validation on unseen synthetic data shows the model can correct in-plane flow topologies for out-of-distribution geometries. Currently, the TTA correction remains strongly localized to the supervised 2D measurement plane, causing volumetric coherence to degrade sharply away from this area and creating disjointed 3D flow structures when the pretrained baseline predictions differ significantly from the ground truth. ... mehrThe application of this framework to in-vitro measurements amplifies non-physical artifacts driven by inherent measurement noise and flexible boundary conditions, underscoring its susceptibility to data outside the training distribution.