Engineers at Purdue University and Children’s National have piloted SMURF, an artificial intelligence framework that rebuilds blood flow through a child’s heart from four-dimensional flow MRI, with no clinician tracing the vessel walls by hand. The results appeared in Medical Image Analysis.
The name stands for Scalable Method for Unsupervised Reconstruction of Flow, and the word unsupervised is the point. The system does not need scans that someone has already labelled. It reads the MRI signal and velocity measurements in tandem, then assembles an anatomy-and-flow model built around the individual patient.
Children are exactly where that matters. Their hearts are smaller, beat faster, and are frequently rebuilt by surgery, all of which leaves supervised tools trained on adult images badly out of their depth. Big labelled datasets of paediatric scans are also close to nonexistent.
Tested on Fontan circulation
Twelve paediatric and young adult cases formed the test set: eight had Fontan circulation, the rerouted plumbing used for single-ventricle hearts, and four had normal anatomy. Human expert tracings and an FDA-cleared clinical processing pipeline served as references.
Its automated segmentations tracked the experts closely, and its flow fields held together more physically than the commercial pipeline’s did. A single graphics processing unit finished each patient in roughly 4.7 to 11.4 minutes.
The bottleneck it removes
4D flow MRI shows how blood swirls and accelerates through the heart, detail that ordinary imaging cannot capture. Analysis has always been the catch: semi-automated workflows ate an hour or more per patient and drifted depending on who did the tracing.
Twelve cases are a pilot, not a validated product, and the authors present the work as a demonstration rather than a cleared tool.
