CHOP Models Children's Hearts in Seconds With Open Source AI

What NVIDIA reports from Children's Hospital of Philadelphia

In a blog post dated September 15, 2026, NVIDIA describes how Children's Hospital of Philadelphia (CHOP) uses open source AI tools to build models of children's hearts in seconds. The stated goal is safer and more precise care for children with congenital heart disease. The post was written by Isha Salian and published on the NVIDIA Blog. It is a vendor blog, so the claims below come from NVIDIA and from the hospital researcher it quotes. We could not check them against an independent source.

The post opens with a simple fact: about 1% of all live births involve a congenital heart defect, and no two are alike. A child may be born with a hole between the lower chambers of the heart, or with a leaking valve in the single pumping chamber that keeps them alive. Such a child needs care that fits their exact anatomy. According to the post, the devices surgeons use were almost never designed with that specific child in mind. Dr. Matthew Jolley, a cardiologist and researcher at CHOP, puts it this way: "You've got a one-of-a-kind kid and an off-the-shelf device." He adds that modeling lets the team find what fits before anyone goes into the cath lab or operating room.

The key facts in the post

CHOP's cardiac modeling service is built on MONAI, an open source medical imaging framework that NVIDIA cofounded. It takes images the care team already has, such as CT scans, MRI and 3D ultrasound, and produces anatomically precise heart models in just seconds. The post says a workflow that once needed four hours of work by a skilled researcher now finishes fast enough for routine clinical use. The post gives some numbers on how far the approach has spread: - More than 20 children's hospitals in the U.S. now run cardiac modeling programs.

• At Boston Children's Hospital, modeling supports more than half of all cardiac surgeries, roughly 500 cases a year.
• CHOP expects to reach about 200 modeled cases this year.
• CHOP now aims to apply the same tools across several disciplines through the IDEA Lab, part of the hospital's Morgan Center for Research and Innovation. The post traces a decade-long path from research to standard care. When Jolley joined CHOP in 2015, 3D echocardiography was just coming online. Tools existed for adult valves, but almost nothing was built for small and complex pediatric anatomy. His lab worked with the open source community to build SlicerHeart, an extension of the 3D Slicer software for visualizing, segmenting and analyzing 3D medical images. The team then developed workflows to model pediatric hearts and valves from several imaging types. For years, one model meant a skilled research assistant spending hours at a workstation. Machine learning changed that. Using MONAI Label and NVIDIA's Auto3DSeg implementation, the team trained segmentation networks on pairs of earlier images and models. NVIDIA says the output meets the quality standard a trained human would produce, in seconds instead of hours. Jolley says machine learning has become "bread and butter": once the team has made 10 or 20 image-model pairs, it trains a model and starts using it.

The post also shows one use case. CHOP can compare how two different artificial valves would fit in a patient's heart, with coloring that shows the stress each valve puts on the cardiac walls. For complex ventricular septal defects, which are holes between the lower chambers, CHOP now models before surgery as a routine step. One early case is described in detail: a child had already had two failed repair attempts, and surgeons could not find the defect with traditional methods. The 3D model clarified the anatomy, and the repair worked on the first try. Jolley says that for cases like these, modeling has moved from research to standard of care.

From shape to physics: Newton and NVIDIA Warp

Seeing the heart is not the end goal. Jolley's team also wants to predict what will happen when a device is placed inside it, before the procedure begins. This is where Newton comes in. The post describes Newton as an open source physics engine, built on the NVIDIA Warp Python framework, that runs physics simulations on GPUs. It was originally meant for simulation-based AI robot training.

CHOP is working with NVIDIA and the open source community on biomechanics-focused simulation frameworks built with Warp that can be brought into Newton. Once linked with 3D Slicer and SlicerHeart, these frameworks could help doctors understand the tissue material properties that decide how a device will deploy in a given patient. With GPU acceleration, the post says, device simulation could fall from up to four hours, or a full overnight run for several configurations, to near real time. A clinician could then compare how different devices fit a child's anatomy and get results fast enough for a same-day decision.

The wording matters here. CHOP has "started" or "begun" to implement features on Warp and Newton for the closure devices that seal holes in children's hearts. It "hopes" to use similar methods for transcatheter valves. The link to SlicerHeart is "being connected", and real-time simulation in clinical workflows is described as a long-term goal. A further tool, a coupler that uses SlicerHeart and NVIDIA Omniverse digital twins powered by OpenUSD, is "in development". The post says such simulations could flow into virtual reality and use embedded vision-language models, so that clinicians can query and interact with a child's cardiac anatomy in simulation before acting on it.

Why open source, according to the post

The post argues that open source suits this field for economic reasons. About 2.4 million people in the U.S. live with congenital heart disease. The post says this group is too rare and too diverse to attract device-company investment at the scale families need, and that no single company has built the tools Jolley's team requires. SlicerHeart's tools are free to use and build on. Researchers at Stanford and Boston Children's add tools alongside CHOP, and a national consortium of children's hospitals is now forming to build the next generation of shared modeling infrastructure.

Jolley sums it up: the population is too small for normal commercial development, but the problem is important enough that the research community and philanthropy are getting behind it. In his words, open source "defies traditional economics" for small and heterogeneous populations, because it allows collaboration without barriers.

The post closes with a point about NVIDIA's own role. It says NVIDIA's investment in open platforms, namely MONAI for medical imaging AI, Newton for physics simulation and OpenUSD for 3D interoperability, gives teams like Jolley's access to infrastructure maintained at industrial scale. A lab at a children's hospital can then use tools that would otherwise need a company-sized engineering team.

Our analysis: what the story shows, and what it does not

Our reading is that the strongest part of this story is the change in workflow, not the AI itself. Segmentation, the step that outlines heart structures in a scan, is slow work when a person does it by hand. The post describes a loop in which a small set of finished image-model pairs trains a network, and the network then does the routine outlining. That is a modest and believable use of machine learning. It removes a bottleneck and leaves the clinical judgment with people.

The second layer, physics simulation of devices, is different in kind. It is the part the post frames in future terms: "can help", "hopes", "in development", "long-term goal". The claim of near real time simulation is a target that depends on integration work that the post says is still under way. Readers should keep the two layers apart. Imaging-based modeling is described as routine at CHOP. Device simulation on Warp and Newton is described as early.

The open source argument is also worth weighing. The economic case is coherent: a rare and varied patient group does not support a normal product market, and shared tools spread the cost. At the same time, the post is published by a company that sells GPUs and platforms, and it presents that company's open projects as the enabling layer. That does not make the claims false, but it does mean the framing favors NVIDIA's products.

Limits of the source and open questions

• The post is self-reported. It gives no accuracy figures, no error rates and no comparison of model outputs with expert outlines. The statement that the output "meets the same quality standard a trained human would produce" is not backed by data in the post.
• The one surgical case, a repair that worked after two failed attempts, is a single example. It shows a possible benefit.

It does not show how often that benefit occurs.

• The post gives no outcome data, such as complication rates or length of stay, for modeled versus non-modeled surgeries.
• Two different "four hours" figures appear. One is the time for a skilled researcher to build a model. The other is the time for a device simulation. They describe different steps.
• The post does not describe any regulatory review, validation study or clinical trial. Whether the simulation tools will need such review before wider clinical use is an open question that the source does not address.
• The post does not say how the consortium of children's hospitals will be funded or governed.

Practical takeaways

For clinicians and hospital innovation teams, the useful lesson is the pattern. Start with images you already have. Build a small set of image-model pairs. Train a segmentation model and use it.

Move to simulation only after the imaging step is trusted. For developers, the post names concrete open projects worth reading: MONAI, MONAI Label, Auto3DSeg, 3D Slicer, SlicerHeart, NVIDIA Warp, Newton and OpenUSD. For policy watchers, the story raises a fair question about how rare-disease tools get built and paid for, and about who validates them. Treat the numbers as NVIDIA's and CHOP's own claims until peer-reviewed or independent results appear.