NV-Generate-CT (rflow-ct)

Purpose

• Used for generating synthetic CT volumes and masks with NV-Generate-CTMR rflow-ct. Not for production training data without review.
• Use the wrapper exactly as documented; do not replace the upstream entrypoint with a handwritten implementation.
• Do not write custom inference code for normal runs. The wrapper owns config staging, output paths, label mapping evidence, and validation.
• Manifest I/O: inputs are configinferoverride; outputs are syntheticctvolumes and result_json.

Instructions

• Read skill_manifest.yaml before changing arguments, side effects, or validation gates.
• Run scripts/runrflowct.py through the documented command below; keep outputs under a caller-provided run directory.
• If a host agent exposes runscript, use runscript("scripts/runrflowct.py", args=[...]); otherwise run the Bash/Python command shown below.
• Emit a single bash code block, and keep the python -m pip install -r "$NVGENERATEROOT/requirements.txt" step in that same command — the runtime may be a fresh environment without nibabel/MONAI, so dropping the install fails with ModuleNotFoundError.
• Do not add rm, mkdir, or any cleanup of --output-dir; the wrapper creates it. Use a fresh --output-dir instead of deleting one.
• Check the emitted JSON and paired verifier guidance before treating the run as evidence.

Available Scripts

Prerequisites

• Required environment variables: NVGENERATEROOT.
• Runtime requirements: GPU/CUDA when declared by the manifest; Python packages listed in runtime.sideeffects.pippackages.
• Side effects: writes generated outputs under the caller's --output-dir, may cache model assets under ~/.cache/huggingface/, and may contact https://huggingface.co or https://github.com during setup.
• Run commands from the repository root unless an existing section below says otherwise.

Limitations

• This is a thin wrapper. Inference, sampling, and decoding are delegated entirely to NVIDIA-Medtech/NV-Generate-CTMR's scripts.inference. Do not modify code under $NVGENERATEROOT.
• rflow-ct requires CUDA and ≈ 16 GB VRAM minimum for the default 256³ outputsize. Larger outputsize (e.g. 512×512×768) needs an A100/H100.
• Output volumes are synthetic. They are not safe to use as training data for production medtech models without an independent quality review.
• Not for clinical deployment, clinical interpretation, autonomous diagnosis, regulatory submission.

Troubleshooting

Wraps the upstream NVIDIA-Medtech/NV-Generate-CTMR rectified-flow synthesis pipeline. The wrapper does not reimplement diffusion, sampling, or autoencoder decoding — it shells out to the upstream scripts.inference entry point exactly as the project's README documents and inspects the produced image/mask pairs.

Preconditions

1. Clone the upstream repo and point NVGENERATEROOT at it (one-time):

``bash test -d "$HOME/nv-generate-ctmr/.git" || \ git clone https://github.com/NVIDIA-Medtech/NV-Generate-CTMR.git $HOME/nv-generate-ctmr export NVGENERATEROOT=$HOME/nv-generate-ctmr pip install -r "$NVGENERATEROOT/requirements.txt" ``

1. Download the rflow-ct weights and the mask-candidate datasets

into the clone (one-time, ≈ 5.5 GB):

``bash cd "$NVGENERATEROOT" python -m scripts.downloadmodeldata --version rflow-ct --root_dir "./" ``

The mask candidates (datasets/allmasksflexiblesizeandspacing4000) condition the diffusion sampler; omitting them via --model_only will make the inference script fail with a missing-file error at startup. The anatomy-size condition file is also part of the full CT download and is needed for controllable mask generation.

1. NVIDIA GPU with ≥ 16 GB VRAM and CUDA. There is no CPU fallback.

For agent-generated user run commands, prefer the short wrapper command in Usage. Do not prepend clone or model-download setup steps when NVGENERATEROOT or the repo-local upstream cache is already present. In a fresh Python environment, still include pip install -r "$NVGENERATEROOT/requirements.txt" before the wrapper unless the active environment has already proven those imports are available; cached weights do not imply cached Python packages. Run the wrapper from the medical-AI-skills repo root. If setup requires cd "$NVGENERATEROOT", return to the Medical AI Skills repo before invoking skills/nv-generate-ct-rflow/scripts/runrflowct.py.

Usage

export NV_GENERATE_ROOT="${NV_GENERATE_ROOT:-$HOME/nv-generate-ctmr}" && \
python -m pip install -r "$NV_GENERATE_ROOT/requirements.txt" && \
python skills/nv-generate-ct-rflow/scripts/run_rflow_ct.py \
  PATH_TO_CONFIG_INFER.json \
  --output-dir runs/nv_generate_ct_rflow_demo \
  --random-seed 0 \
  --version rflow-ct

Replace PATHTOCONFIGINFER.json with the user's actual request/config path. Do not copy the fixture path from this document unless the user explicitly asked to run that fixture. If the user says "the case request is at runs/.../chestlungtumorcontrollable.json", that exact path is the first positional argument to scripts/runrflowct.py.

The fixture argument is a configinfer.json override file: it can replace numoutputsamples, bodyregion, anatomylist, controllableanatomysize, outputsize, and spacing. Pass default to use the upstream config verbatim. The wrapper stages the override into the upstream tree before running.

Fixture catalog

fixtures/ ships curated configs for common paired synthesis use cases: chest lung lobes, chest with controllable lung tumor, abdomen solid organs, abdomen with controllable hepatic tumor, head + cervical spine, pelvis. See [fixtures/README.md](fixtures/README.md) for the full table.

Helper commands

# Browse the 132-class label_dict grouped by body region.
python skills/nv-generate-ct-rflow/scripts/list_anatomies.py --region chest
python skills/nv-generate-ct-rflow/scripts/list_anatomies.py --controllable
python skills/nv-generate-ct-rflow/scripts/list_anatomies.py --filter tumor

# Validate a fixture and preview cost without launching inference.
NV_GENERATE_ROOT=$HOME/nv-generate-ctmr \
  python skills/nv-generate-ct-rflow/scripts/run_rflow_ct.py \
    skills/nv-generate-ct-rflow/fixtures/abdomen_liver_spleen.json \
    --output-dir runs/preview --preflight-only

Advanced helpers stay inside this skill for debugging and less-common CT generation modes. Use them only when the user explicitly asks for that mode:

# Raw MAISI mask diagnostic, useful for checking lung tumor -> label 23.
python skills/nv-generate-ct-rflow/scripts/run_ct_mask.py \
  skills/nv-generate-ct-rflow/fixtures/ct_mask_lung_tumor.json \
  --output-dir runs/ct_mask_debug --preflight-only

# CT image from an existing MAISI label mask with body label 200.
python skills/nv-generate-ct-rflow/scripts/run_ct_from_mask.py \
  skills/nv-generate-ct-rflow/fixtures/ct_from_mask_request_example.json \
  --output-dir runs/ct_from_mask_demo

# CT image-only generation without paired labels.
python skills/nv-generate-ct-rflow/scripts/run_ct_image.py \
  skills/nv-generate-ct-rflow/fixtures/ct_image_only_default.json \
  --output-dir runs/ct_image_only_demo --version rflow-ct

The wrapper runs preflight on every invocation (regardless of --preflight-only): config-schema bounds, anatomy names matched against the upstream labeldict, bodyregion in the supported set, controllableanatomysize constraints, upstream CT output-size/spacing contracts, body-region-aware x/y FOV minimums, dataset presence under $NVGENERATEROOT/datasets/, CUDA available, and an estimated peak VRAM / wall-time. Runs estimated to exceed 5 min wall-time or 30 GB VRAM peak require --yes to proceed.

Each invocation runs python -m scripts.inference -t configs/confignetworkrflow.json -i configs/configinfer.json -e configs/environmentrflow-ct.json --random-seed <s> --version rflow-ct. Output evidence records the upstream git commit, model checkpoint hashes, the rendered config, per-sample image/mask geometry, mask label set, image HU range summary, and per-class voxel volumes.

When controllableanatomysize is non-empty, upstream ignores the broader anatomylist for the saved paired label map and filters labels to the controllable anatomy names. The saved paired label values are local 1..N ordinals, not raw MAISI label IDs. Read output.outputlabelmapping in resultjson to map saved output labels back to source labels; for example, output label 1 can represent MAISI label 23 (lung tumor). For curated lung-tumor examples, prefer a controllable size around 0.5 or larger; smaller requests such as 0.2 can produce absent or extremely small label-23 components for some seeds.

For FOV and setup details, see references/fov-and-downloads.md. For advanced helper label-space details, see references/ct-mask-label-space.md and references/ct-from-mask-format.md.

Visual sample card

Alongside the NIfTI pairs, the wrapper writes summary.html to the output directory: a per-sample mid-slice triptych (axial / coronal / sagittal) with label overlay, plus a table of the rendered config and verifier-facing aggregates. Lets you eyeball the result without firing up 3D Slicer. Pass --no-summary-card to skip.

Anatomy plausibility (label-set sanity, voxel HU range as CT, image/mask geometry match, declared output labels present, lung-lobe HU floor) is checked by verifiers/ctsynthesisquality_v1.

Not for clinical interpretation, training data for production deployment, or any non-synthetic-research use.