MedTech
Open Models
Foundation models for 3D medical image segmentation, synthetic data generation, clinical reasoning, and physical AI — now runnable by your coding agent.
Physical AI
3 models built on Cosmos and GR00T for vision-language-action control, surgical world modeling, and sim-to-real transfer via Isaac and Holoscan
GR00T-H
Healthcare Physical AI VLA
Vision-language-action model for surgical robot control, post-trained on 601 hours of surgical data from 7 robot platforms.
Model informationCosmos-H-Surgical
Surgical World Model
Predict future surgical video from an image and text, and turn simulation control videos into photorealistic surgical video.
Model informationCosmos-H-Surgical-Simulator
Action-Conditioned Simulation
Generate surgical video from robot kinematic actions for policy evaluation in simulation and synthetic data generation.
Model informationMedical Imaging
9 models spanning MONAI-powered segmentation, raw signal reconstruction, synthetic volume generation, and clinical VLM reasoning across CT, MRI, X-ray, and ultrasound
NV-Segment-CT
132-Class CT Segmentation
Automatic and interactive segmentation of 132 anatomical structures in CT images with point-click refinement.
Model informationNV-Segment-CTMR
345+ Class CT & MRI Segmentation
Automatic segmentation of 345+ anatomical structures across CT body, MRI body, and MRI brain modalities.
Model informationNV-Raw2Insights-MRI
Raw MRI Signal Reconstruction
Reconstruct MR images from undersampled k-space in a single forward pass at 8x–24x acceleration; trained on cardiac MRI.
Model informationNV-Raw2Insights-US
Raw Ultrasound Processing
Estimate speed-of-sound maps from raw ultrasound channel data so beamforming can correct for tissue and sharpen images.
Model informationNV-Generate-CT
3D CT Synthesis
Generate high-resolution synthetic CT volumes up to 512x512x768 with controllable anatomy and paired segmentation masks.
Model informationNV-Generate-MR
3D MRI Synthesis
Generate synthetic 3D MRI volumes up to 512x512x768 by body region, with optional masks for up to 127 anatomical classes.
Model informationNV-Generate-MR-Brain
3D Brain MRI Synthesis
Generate synthetic brain MRI across T1, T2, FLAIR, and SWI contrasts at up to 512x512x256 and 0.45 mm in-plane resolution.
Model informationNV-Reason-CXR
Chest X-Ray Clinical VLM
Vision-language model for chest X-ray interpretation with step-by-step chain-of-thought clinical reasoning.
Model informationNV-Reason-CT
3D CT Reasoning VLM
Vision-language model for native 3D chest and abdominal CT with structured reports, visual QA, and step-by-step reasoning.
Model informationRun these models with your coding agent
14 agent skills cover the imaging models — a run skill for the segmentation, generation, and reasoning models, finetune skills, and 4 DICOM utilities. Install one and Claude Code, Cursor, or Codex runs the model with the right entrypoint, flags, and inputs. Engineering and research use only.
Every skill ships its validation
- Benchmark report
- Eval prompts
- Tests
- Output validators
Each model page links them, so you can review the evidence yourself — engineering reproducibility, not a clinical claim. How skills are validated.
npx skills add NVIDIA-Medtech/medical-AI-skillsOpen the Agent Skills spec at agentskills.io, or browse all on GitHub.
Run a model
7Finetune a model
3Why MedTech Open Models
Built for medical imaging researchers and developers
Start from Strong Foundations
Foundation models pre-trained on large-scale medical datasets. Fine-tune on your data instead of training from scratch.
From 2D to 4D
CT, MRI, X-ray, ultrasound, and surgical video in one collection. Static images through dynamic physiology.
Synthetic Data Generation
Create synthetic volumes to augment training, simulate rare pathologies, and enable privacy-preserving data sharing across institutions.
Open Weights
Download from HuggingFace and run locally. No API keys, no cloud dependency. Apache 2.0 source with model weights for commercial and research use.
Run with Your Agent
Agent skills wrap each model so Claude Code, Cursor, or Codex calls it correctly — right flags, right inputs — from a plain-language prompt.
Research to Deployment
Every model includes pre-trained weights, example configs, and documentation. Go from paper to inference with minimal setup.