NV-Segment-CTMR

Purpose

• Used for running NV-Segment-CTMR on CT or MRI NIfTI volumes and recording label-map evidence. Not for clinical interpretation.
• Use the wrapper exactly as documented; do not replace the upstream entrypoint with a handwritten implementation.
• Manifest I/O: inputs are ctormrvolume; outputs are labelmap and result_json.

Instructions

• Read skill_manifest.yaml before changing arguments, side effects, or validation gates.
• Run scripts/run_ctmr.py through the documented command below; keep outputs under a caller-provided run directory.
• If a host agent exposes runscript, use runscript("scripts/run_ctmr.py", args=[...]); otherwise run the Bash/Python command shown below.
• Check the emitted JSON and paired verifier guidance before treating the run as evidence.

Available Scripts

Prerequisites

• Runtime requirements: GPU/CUDA when declared by the manifest; Python packages listed in runtime.sideeffects.pippackages.
• Optional environment variables: NVSEGMENTCTMRROOT selects the trusted upstream checkout; CUDAVISIBLEDEVICES restricts visible GPUs; MONAIDATADIRECTORY and PYTORCHCUDAALLOCCONF override the wrapper's output-local cache and allocator defaults when needed.
• Side effects: writes segmentation outputs under the caller's --output-dir, may cache model assets under ~/.cache/huggingface/, and may contact https://github.com or https://huggingface.co during setup.
• Run commands from the repository root unless an existing section below says otherwise.

Limitations

• This is a thin wrapper. Inference, preprocessing, and postprocessing are delegated entirely to the upstream MONAI bundle under $NVSEGMENTCTMRROOT or the repo-local fallback at .workbenchdata/upstreams/NV-Segment-CTMR/NV-Segment-CTMR.
• The default wrapper path runs automatic "segment everything" inference for CTBODY, MRIBODY, or MRIBRAIN. MRIBRAIN inputs must already follow the upstream brain preprocessing requirements.
• Label names are loaded from upstream configs when available. If a label dictionary is absent, the wrapper still records label IDs and marks only negative IDs as invalid.
• No clinical, diagnostic, regulatory, or treatment-planning claims.
• Not for clinical deployment, clinical interpretation, autonomous diagnosis, regulatory submission.

Troubleshooting

Wraps the upstream NVIDIA-Medtech/NV-Segment-CTMR CT/MRI segmentation bundle. The wrapper does not reimplement VISTA3D inference. It shells out to the documented python -m monai.bundle run entry point, then inspects the produced NIfTI label map.

Exact Runnable Surface

For CT body segmentation user runs and benchmark answers, use this fresh-environment-safe repo-root command shape exactly:

export NV_SEGMENT_CTMR_ROOT="${NV_SEGMENT_CTMR_ROOT:-$HOME/.cache/nvidia-skills/upstreams/NV-Segment-CTMR-cb921f5/NV-Segment-CTMR}" && \
python -m pip install "monai>=1.5,<1.6" "numpy<2" nibabel scipy typer PyYAML fire huggingface_hub pytorch-ignite einops && \
python skills/nv-segment-ctmr/scripts/run_ctmr.py PATH_TO_IMAGE.nii.gz --modality CT_BODY --output-dir OUT_DIR

Do not invent python -m nvsegmentctmr, infer.py, or Medical AI Skills run commands. PATHTOIMAGE.nii.gz must be the user's supplied input path. For benchmark/user run answers, the bash block is invalid if it includes mkdir -p .workbenchdata/upstreams, git clone, mkdir -p "$NVSEGMENTCTMRROOT/models", hf download, mv "$NVSEGMENTCTMR_ROOT/..., or any other command that creates, downloads into, or moves files inside the shared upstream checkout.

Preconditions

One-time maintainer setup only; do not include these commands in user answers or benchmark commands. The benchmark environment already provides the repo-local upstream cache and model files.

If NVSEGMENTCTMR_ROOT already names a local bundle checkout, the wrapper uses it and records its current commit in the result. Otherwise, clone the recommended pinned default once:

if [ -z "${NV_SEGMENT_CTMR_ROOT:-}" ]; then
  export NV_SEGMENT_CTMR_COMMIT=cb921f5c58837c0f42a713855d68b32af88e1cdd
  export NV_SEGMENT_CTMR_CHECKOUT="$HOME/.cache/nvidia-skills/upstreams/NV-Segment-CTMR-cb921f5"
  if [ ! -d "$NV_SEGMENT_CTMR_CHECKOUT/.git" ]; then
    git clone https://github.com/NVIDIA-Medtech/NV-Segment-CTMR.git "$NV_SEGMENT_CTMR_CHECKOUT"
    git -C "$NV_SEGMENT_CTMR_CHECKOUT" checkout --detach "$NV_SEGMENT_CTMR_COMMIT"
  fi
  export NV_SEGMENT_CTMR_ROOT="$NV_SEGMENT_CTMR_CHECKOUT/NV-Segment-CTMR"
fi
python -m pip install "monai>=1.5,<1.6" "numpy<2" nibabel scipy typer PyYAML fire huggingface_hub pytorch-ignite einops && \
python -c "import monai, nibabel, numpy"

mkdir -p "$NV_SEGMENT_CTMR_ROOT/models"
test -e "$NV_SEGMENT_CTMR_ROOT/models/model.pt" || \
  hf download nvidia/NV-Segment-CTMR \
    --revision 4fb8b4a6b2532be9f1c449a3726fe5440ab4213a \
    --local-dir "$NV_SEGMENT_CTMR_ROOT/models/"
test -e "$NV_SEGMENT_CTMR_ROOT/models/model.pt" || \
  mv "$NV_SEGMENT_CTMR_ROOT/models/vista3d_pretrained_model/model.pt" \
    "$NV_SEGMENT_CTMR_ROOT/models/model.pt"

The wrapper also searches .workbenchdata/upstreams/NV-Segment-CTMR/NV-Segment-CTMR if NVSEGMENTCTMRROOT is unset or does not have the required bundle layout.

For agent-generated user run commands, use the command in Usage. Do not copy the one-time Preconditions block into the answer: do not create or write under $NVSEGMENTCTMRROOT, do not run hf download, and do not move files in the shared upstream checkout during a benchmark or user run. Do not prepend pip install -r "$NVSEGMENTCTMRROOT/requirements.txt" in a Python 3.12 environment; the upstream requirements pin NumPy 1.24.4, which does not build cleanly there. In a fresh Python environment, install the minimal compatible runtime shown above (monai>=1.5,<1.6, numpy<2, nibabel, scipy, typer, PyYAML, fire, huggingface_hub, pytorch-ignite, einops) before the wrapper. Cached models do not imply cached Python packages.

Runtime needs an NVIDIA GPU with CUDA. The upstream bundle may import on CPU-only hosts, but this skill is declared as CUDA-required because the published workflow is a 3D CT/MRI foundation model inference path.

Usage

From Medical AI Skills repo root:

export NV_SEGMENT_CTMR_ROOT="${NV_SEGMENT_CTMR_ROOT:-$HOME/.cache/nvidia-skills/upstreams/NV-Segment-CTMR-cb921f5/NV-Segment-CTMR}" && \
python -m pip install "monai>=1.5,<1.6" "numpy<2" nibabel scipy typer PyYAML fire huggingface_hub pytorch-ignite einops && \
python skills/nv-segment-ctmr/scripts/run_ctmr.py PATH_TO_IMAGE.nii.gz \
  --modality CT_BODY \
  --output-dir runs/nv_segment_ctmr_demo

Replace PATHTOIMAGE.nii.gz with the user's actual input path. Do not copy the example fixture path into a user run. If the user provides an explicit input path under runs/, that path must be the first positional argument to scripts/run_ctmr.py.

Supported automatic segmentation modalities are CTBODY, MRIBODY, and MRIBRAIN. For MRIBRAIN, the upstream README requires brain-specific preprocessing before bundle inference; pass an already preprocessed image to this wrapper.

Pass --label-prompts "3,14" to request specific upstream class IDs instead of only the modality-level "segment everything" set. The evidence output records input geometry, output mask path, observed label IDs, unexpected labels, per-class voxel counts, per-class physical volumes from the mask header spacing, runtime, upstream command, model inventory, and geometry checks.

Pass --ground-truth PATH to record a reference label-map path under input.groundtruthpath. The skill does not compute Dice; that is the paired verifier's job.

Anatomy plausibility and optional per-class Dice/IoU against the recorded ground truth can be checked by verifiers/ctsegmentationquality_v1 for CT-body outputs.

Not for clinical interpretation, production deployment, autonomous diagnosis, or regulatory submission.