Used for generating synthetic CT volumes and masks with NV-Generate-CTMR rflow-ct. Not for production training data without review.
This NVIDIA MedTech skill is a thin, carefully-scoped wrapper for generating synthetic CT volumes and anatomy masks using the NV-Generate-CTMR rflow-ct model. It solves the problem of running a complex upstream medical-image generation pipeline reproducibly: the wrapper owns config staging, output paths, label-mapping evidence, and validation, and the skill repeatedly instructs the agent not to replace the upstream entrypoint with handwritten inference code. Inference, sampling, and decoding are delegated entirely to the upstream repository.
The primary entrypoint, `run_rflow_ct.py`, takes a config-override JSON and a caller-provided output directory and produces synthetic CT volumes plus a result JSON. A curated fixture catalog covers common anatomy scenarios (chest lung lobes and tumors, abdomen organs and hepatic tumor, head/brain, and pelvis), and a `list_anatomies.py` helper enumerates supported anatomies and regions. Additional advanced helpers generate raw MAISI masks, CT from an existing mask, or image-only CT. The skill documents a `controllable_anatomy_size` convention validated by preflight checks, an allow-listed set of override keys that rejects unknown keys, and prerequisites including a `NV_GENERATE_ROOT` environment variable, CUDA/GPU with roughly 16 GB VRAM, and a pip install from the upstream requirements. Documented side effects are limited to writing outputs under the caller's output directory, caching model assets under the Hugging Face cache, and contacting huggingface.co or github.com during setup. It explicitly forbids adding `rm`/cleanup commands.
Target users are medical-imaging researchers and ML engineers generating synthetic training data. The skill states firmly that outputs are synthetic, are not safe as production training data without independent quality review, and are not for clinical deployment, interpretation, autonomous diagnosis, or regulatory submission.
Synthetic CT volumes and anatomy masks generated by the NV-Generate-CTMR rflow-ct model, plus a result JSON, written under a caller-provided output directory.
It requires the `NV_GENERATE_ROOT` environment variable, a CUDA GPU with roughly 16 GB VRAM minimum for the default 256-cubed output (larger outputs need an A100/H100), and a pip install of the upstream requirements since the runtime may be a fresh environment.
The skill is explicit that outputs are synthetic and not safe as production training data without an independent quality review, and are not for clinical deployment, interpretation, autonomous diagnosis, or regulatory submission.
It explicitly forbids adding `rm`, `mkdir`, or cleanup of the output directory. Its network access is limited to contacting huggingface.co and github.com during setup and caching model assets under the Hugging Face cache.
No. The skill instructs using the documented wrapper exactly and delegating all inference, sampling, and decoding to the upstream repository; it says not to modify upstream code.
Quick Setup:
.claude/skills/Repository
nvidia/skills