CHOP turns children’s heart scans into patient-specific models in seconds

Children’s Hospital of Philadelphia is using open-source medical-imaging AI to turn children’s cardiac scans into patient-specific 3D models in seconds. The practical leap is speed: NVIDIA says work that once occupied a skilled researcher for four hours can now fit into routine clinical care.

NVIDIA Watch analysis

What happened

CHOP’s cardiac-modelling service takes existing CT, MRI and 3D ultrasound images and segments the relevant anatomy using networks trained with MONAI Label and NVIDIA’s Auto3DSeg implementation. The resulting models help clinicians inspect complex congenital heart defects before entering the catheterisation laboratory or operating theatre.

According to NVIDIA’s account of the programme, more than 20 US children’s hospitals now run cardiac-modelling services. Boston Children’s reportedly uses modelling in more than half of its cardiac surgeries, around 500 cases annually, while CHOP expects about 200 modelled cases this year.

Key findings

• Four hours compressed into seconds
  Machine-learning segmentation has made a previously labour-intensive modelling workflow fast enough for routine use, according to NVIDIA.
• Existing scans provide the input
  The workflow works from CT, MRI and 3D ultrasound images already collected by a child’s clinical team.
• Open source fills a narrow market gap
  SlicerHeart and MONAI let hospitals collaborate on tools for varied paediatric anatomies that commercial developers may struggle to serve economically.
• Simulation is the next step
  CHOP is developing GPU-powered biomechanics tools to test how closure devices or valves might behave in a patient-specific heart model.
• Clinical modelling and device simulation differ in maturity
  Cardiac modelling is already in use, while near-real-time device comparison remains a development goal rather than an established clinical capability.

Why it matters

Congenital heart defects vary enormously, while surgeons often work with devices that were not designed for one child’s exact anatomy. Faster modelling gives clinical teams more opportunity to inspect that anatomy before a procedure without turning every case into a small research project.

The more ambitious prospect is simulation. CHOP is working with NVIDIA and open-source contributors to connect SlicerHeart with the Warp and Newton physics frameworks, aiming to reduce device simulations from hours or overnight runs to something useful for same-day decisions. That could let clinicians compare configurations before acting, but the report does not provide independent clinical validation of those simulations.

Our read

This is a compelling example of medical AI doing a specific, comprehensible job: converting scans into usable anatomical models quickly enough to change a workflow. No diagnostic oracle, no robot surgeon taking the lift to theatre, just a difficult modelling task becoming dramatically faster.

The important boundary is between what works now and what comes next. The segmentation workflow appears operational; near-real-time device simulation is still being assembled. Hospitals assessing similar systems should ask for validation across scanners, anatomies and clinical teams, alongside evidence that faster modelling improves decisions and outcomes.

What to watch

• Independent clinical validation of model accuracy and patient outcomes.
• Whether near-real-time device simulation reaches routine clinical use.
• Expansion of CHOP’s workflow beyond paediatric cardiac care.
• How the emerging hospital consortium governs shared models, data and software.

Discussion spark: Which evidence should hospitals require before patient-specific AI models or simulations become part of routine procedural decisions?

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.