This project develops SMURF, a patient-specific framework that infers a probabilistic geometry field and a continuous velocity field directly from 4D flow MRI magnitude and phase-derived velocity data. The method formulates a coupled likelihood where shared class probabilities connect the geometry and velocity terms, so segmentation and velocity reconstruction are solved as one inference problem.
The fields are represented as coordinate-based models that can be evaluated on the acquisition grid or a finer query grid. Training uses subsampling over space and time to control compute cost and runs on a single GPU per case. Outputs include time-resolved segmentations, reconstructed velocity fields, and derived diagnostics such as conservation residuals computed from the reconstructed field.
