Patient-specific unsupervised neural reconstruction of cardiac blood flow from 4D flow MRI
TL;DR
SMURF (Scalable Method for Unsupervised Reconstruction of Flow), a label-free framework that infers probabilistic geometry and velocity as implicit neural fields from magnitude and phase data through a measurement model, reduces reliance on expert-drawn segmentations and produces time-resolved segmentations and reconstructed flow fields that are more physically consistent from 4D flow MRI alone.
Abstract
Clinical 4D flow MRI can provide detailed measurements of cardiac blood flow, but current workflows rely on manual segmentation and post hoc velocity filtering, which introduce variability and reduce physical consistency. We introduce SMURF (Scalable Method for Unsupervised Reconstruction of Flow), a label-free framework that infers probabilistic geometry and velocity as implicit neural fields from magnitude and phase data through a measurement model. This coupling casts segmentation and velocity reconstruction as a single inference problem and enables evaluation on finer grids without retraining. In 12 pediatric cases (four Normal, eight Fontan), SMURF matches expert segmentations with surface offsets of ∼1 voxel (Normal) and ∼1.3 voxels (Fontan), reduces RMS divergence and vorticity-transport momentum residuals by roughly 70%-80% relative to an FDA-cleared post-processing pipeline, and completes time-resolved segmentation and velocity reconstruction in 4.7-11.4 min per case, about five- to fifteen-fold shorter than reported semiautomatic post-processing. SMURF reduces reliance on expert-drawn segmentations and produces time-resolved segmentations and reconstructed flow fields that are more physically consistent from 4D flow MRI alone.