Open foundation models for physical AI and medical imaging.
- 11 models for physical AI, segmentation, generation, and reasoning
- 5 modalities: surgical video, CT, MRI, X-ray, and ultrasound
- 12 agent skills to run the models from your coding agent:
npx skills add NVIDIA-Medtech/medical-AI-skills - Apache 2.0 source code. Model weights available on HuggingFace
We have also released Open-H, a community-wide initiative for physical AI datasets for healthcare robotics. Browse the dataset on HuggingFace.
NVIDIA Isaac for Healthcare is the three-computer solution for healthcare robotics: a domain-specific framework on Isaac Sim, Isaac Lab, Omniverse, and Holoscan for simulating, training, and deploying AI-enabled medical robots across surgical robotics, diagnostic imaging (ultrasound, CT, CBCT), and hospital and patient care.
Explore Isaac for Healthcare →
Physical AI for surgical simulation and embodied robot control, built on NVIDIA Cosmos and Isaac, deploys on Holoscan.
- GR00T-H: Healthcare variant of GR00T-N, post-trained for multi-embodiment surgical robot control across 16 platforms
- Cosmos-H-Surgical: Surgical video world model with Predict (future-state generation) and Transfer (control-conditioned generation)
- Cosmos-H-Surgical-Simulator: Action-conditioned video generation for in-silico policy evaluation across 9 surgical robot embodiments
| Model | Task | Modality | License | Links |
|---|---|---|---|---|
| GR00T-H | Surgical robot VLA | Multi | Non-Commercial | Repo · HF |
| Cosmos-H-Surgical | Surgical world model | Video | Non-Commercial | Repo · HF · Paper |
| Cosmos-H-Surgical-Sim | Action-conditioned sim | Video | Commercial | Repo · HF |
Foundation models for 3D segmentation, synthetic volume generation, signal reconstruction, and clinical reasoning covering CT, MRI, X-ray, and ultrasound.
- NV-Segment-CT: Segment with Automatic and interactive (point-click) methods for 132 anatomical structures in CT
- NV-Segment-CTMR: Segment 345+ classes across CT body, MRI body, and MRI brain
- NV-Raw2Insights-MRI: Transforms raw k-space signals to high-quality MRI with accelerated reconstruction and motion correction
- NV-Raw2Insights-US: Transforms raw ultrasound measurements to clinical-quality images using physics-informed AI
- NV-Generate-CT: Paired CT images and segmentation masks up to 512x512x768
- NV-Generate-MR: Rectified Flow across brain, abdomen, breast, and prostate MRI
- NV-Generate-MR-Brain: Brain volumes across T1w, T2w, FLAIR, and SWI contrasts with cross-sequence synthesis via ControlNet
- Live demo
- NV-Reason-CXR: 3B-parameter reasoning VLM built on Qwen2.5-VL that utilizes chain-of-thought reasoning for chest X-rays.
- Live demo
| Model | Task | Modality | License | Links |
|---|---|---|---|---|
| NV-Segment-CT | 132-class segmentation | CT | Commercial | Repo · HF · Paper |
| NV-Segment-CTMR | 345+ class segmentation | CT, MRI | Non-Commercial | Repo · HF · Paper |
| NV-Raw2Insights-MRI | Signal reconstruction | MRI | Commercial | Repo · HF |
| NV-Raw2Insights-US | Signal reconstruction | Ultrasound | Non-Commercial | Repo · HF |
| NV-Generate-CT | 3D volume synthesis | CT | Commercial | Repo · HF · Paper |
| NV-Generate-MR | 3D volume synthesis | MRI | Non-Commercial | Repo · HF · Paper |
| NV-Generate-MR-Brain | Brain MRI synthesis | MRI | Commercial | Repo · HF · Paper |
| NV-Reason-CXR | Chest X-ray reasoning | X-Ray | Non-Commercial | Repo · HF · Paper |
Run the models from your coding agent: install a skill and Claude Code, Cursor, or Codex calls the model correctly on the first try, from a plain-language prompt. Skills follow the open Agent Skills spec.
npx skills add NVIDIA-Medtech/medical-AI-skills- Run: NV-Segment-CT, NV-Segment-CTMR, NV-Generate-CT, NV-Generate-MR, NV-Generate-MR-Brain, NV-Reason-CXR
- Finetune: NV-Segment-CT, NV-Generate-CT (MAISI VAE), NV-Generate-MR-Brain
- DICOM utilities: preflight, metadata extraction, and DICOM→NIfTI conversion
Models are released under the NVIDIA Open Model License (commercial) or NVIDIA Non-Commercial License. Source code is Apache 2.0. Check each model's repository for the specific license terms.
