SMURF AI maps 4D flow MRI images of complex blood flow in CHD - Children's National

Researchers from Children's National and Purdue University successfully piloted an artificial intelligence (AI) framework called the Scalable Method for Unsupervised Reconstruction of Flow (SMURF) that reconstructs complex congenital heart disease (CHD) and blood-flow patterns from four-dimensional (4D) flow magnetic resonance imaging without the need for hands-on manual segmentation by clinical experts. The findings were published in Medical Image Analysis.

What it means

In the study, authors note that pediatric cardiac imaging presents unique challenges because of smaller anatomy, faster heart rates and highly variable post-surgical anatomy. These factors make traditional image processing more difficult and limit the availability of large training datasets for supervised AI approaches.

Rather than relying on pre-labeled training datasets, SMURF uses an unsupervised machine learning approach to analyze magnetic resonance imaging (MRI) and velocity data together, creating a patient-specific model for cardiac anatomy and blood flow.

The investigators tested SMURF using 12 pediatric and young adult clinical cases, including four patients with normal cardiac anatomy and eight patients with Fontan circulation. Expert-generated segmentations and a clinically deployed FDA-cleared processing pipeline served as reference standards.

Results showed that SMURF automatically generated segmentations closely matched with human expert assessments. Reconstructed blood-flow fields demonstrated substantially greater physical consistency than those produced by the clinical comparison pipeline. The approach was particularly effective at mapping flow and structure in patients who have undergone a Fontan procedure for single ventricle heart defects.

Why it matters

4D flow MRI provides detailed information about how blood moves through the heart and blood vessels. The tremendous clinical value has often been offset by the fact that generating meaningful measurements required manual or semi-automated image processing, which is labor-intensive and often subjective.

The authors note that processing time reductions were among the most notable findings. SMURF completed fully automated, time-resolved segmentation and flow reconstruction in approximately 4.7 to 11.4 minutes per patient when run on a single graphics processing unit. Previously reported semi-automated 4D flow MRI workflows required an hour or more.

This study demonstrates that a patient-specific AI approach can successfully reconstruct cardiac anatomy and blood flow based on 4D flow MRI. This may eliminate the need for manual segmentation and allow for faster and more automated assessment of cardiovascular function in children with complex cardiac conditions. If so, this technology could allow for greater adoption of 4D flow MRI as a tool for clinical assessment and surgical planning in CHD.

"Traditional 4D flow MRI analysis treats anatomy and blood-flow reconstruction as separate problems," the researchers explain. "SMURF brings them together into a single framework, allowing information from blood flow and anatomy to improve one another simultaneously."

Children's National leads the way

Future work with SMURF will focus on expanding validation in larger patient populations and incorporating additional hemodynamic measurements, including pressure and energy-related metrics. The authors also note that the probabilistic nature of the framework could potentially support uncertainty estimation and improved quantification of clinically important biomarkers.

Read the study, Patient-specific unsupervised neural reconstruction of cardiac blood flow from 4D flow MRI published in Medical Image Analysis.