Philadelphia children's hospital models hearts in seconds with Nvidia open-source code

Children's Hospital of Philadelphia has moved pediatric heart modeling from research into production: a pipeline that once required four hours of a skilled researcher now runs in seconds, thanks to segmentation models built on MONAI, the open-source medical-imaging framework founded by Nvidia. The result is an anatomically accurate model available before every catheterization or surgery for children whose heart defects are one of a kind.

The problem: a unique child, an off-the-shelf device

Roughly one percent of live births involve a congenital heart defect, and no two are alike. "You have a unique child and an off-the-shelf device," explains Dr. Matthew Jolley, a cardiologist and researcher at CHOP. Historically, the implants and tools surgeons reach for were designed for standard adult anatomy — not for a single ventricle with a leaky valve or a perforated septum between the lower chambers. That gap forced teams to improvise in real time in the operating room.

The solution: from MONAI Label to Auto3DSeg

Jolley's lab built SlicerHeart, an extension of 3D Slicer for visualization, segmentation and analysis of 3D medical images, and trained segmentation networks on prior image-model pairs using MONAI Label and Nvidia's Auto3DSeg implementation. After 10 to 20 training pairs, the automation reaches the quality of a human annotator. "Machine learning has become bread and butter," Jolley says; once there is enough data, the model enters routine use.

Broad adoption: more than 20 hospitals, Boston leading

More than 20 children's hospitals across the United States now run cardiac modeling programs. At Boston Children's, modeling supports more than half of all heart surgeries — roughly 500 cases a year. CHOP expects about 200 modeled cases this year. The work is centered in the IDEA Lab, part of the hospital's Morgan Center for Research and Innovation, with the goal of extending the same tools to other disciplines.

Clinical impact: from failed repair to first-attempt success

The clinical turn is clearest in complex ventricular septal defects. In one early case, a child underwent two failed repair surgeries because surgeons could not locate the defect with traditional methods. The 3D model clarified the anatomy, and the repair succeeded on the first attempt. For such cases, Jolley says, modeling has become the standard of care, not a pilot.

Next step: Newton, Warp and biomechanical simulation

Visualization alone is not enough. The team wants to predict how tissue will respond when a device is deployed inside it, before the catheter enters. That is where Newton comes in — an open-source physics engine built on Warp, Nvidia's Python framework for GPU-accelerated simulation. CHOP is collaborating with Nvidia and the broader community to build biomechanical simulation frameworks that will integrate into 3D Slicer and SlicerHeart, letting clinicians test virtual tissue behavior before every procedure.