NV-Segment-CTMR Medical Image Segmentation
Run CT or MRI NIfTI segmentation and capture inspectable label-map evidence.
The skill clearly limits use to engineering verification, discloses output, cache, GitHub/Hugging Face network access, GPU, and environment modification, and prohibits clinical interpretation. It does not require execution confirmation, defaults to modifying the active Python environment, may process sensitive medical images, and lacks explicit privacy handling, isolation, rollback, and output-recovery controls, so points are deducted.
Entrypoints, arguments, dependencies, failure cases, timeouts, and several unit tests are documented; tests cover command construction, output discovery, labels, and geometry. However, this review did not execute key paths, external bundle/model/GPU dependencies are not reproducible from the files alone, and the wrapper exits with code 0 even when the subprocess reports failure, so points are deducted and the static ceiling applies.
The audience, NIfTI input, CT_BODY/MRI_BODY/MRI_BRAIN modalities, outputs, and non-clinical boundary are fairly clear, including the MRI_BRAIN preprocessing caveat. Trigger and non-fit coverage remain limited, Chinese-language guidance is absent, and core setup depends on GitHub/Hugging Face, which may be difficult to reach from mainland China, so points are deducted.
SKILL.md provides purpose, instructions, prerequisites, limitations, troubleshooting, and usage; the manifest fixes version, dependencies, I/O, upstream revisions, and a verifier, while licensing is stated. However, the recommended Examples section is missing, maintenance ownership and changelog practices are not specific, and setup/runtime constraints contain hidden complexity, so points are deducted.
The skill can invoke segmentation and produce directly structured JSON containing label-map, geometry, volume, and runtime evidence, while explicitly excluding Dice computation and clinical quality claims. The committed benchmark reports two passing tasks and some correctness/effectiveness uplift, but coverage is small, there are no negative tasks, and no independent reproduction evidence is available, so points are deducted.
Pinned upstream commits, a model revision, output schema, implementation, unit tests, and a benchmark report provide auditable material. This review did not execute the tests, and the benchmark is only an aggregate report without raw run artifacts or independent corroboration, so only limited static evidence credit is justified.
- Do not use outputs for diagnosis, treatment, clinical deployment, or regulatory conclusions; MRI_BRAIN requires the upstream brain-specific preprocessing.
- Run in an isolated environment after confirming network access, model availability, GPU capacity, and permission to modify the active Python environment; do not send patient data, credentials, or sensitive logs to external services.
- The subprocess-failure exit-code behavior should be corrected, and raw execution artifacts, negative trigger tests, and broader modality coverage should be added.
What does this skill do, and when should you use it?
NV-Segment-CTMR is an NVIDIA Medical AI Skill for segmenting CT and MRI NIfTI volumes. It uses a documented wrapper to invoke the upstream MONAI NV-Segment-CTMR bundle and supports CT_BODY, MRI_BODY, and MRI_BRAIN automatic segmentation. The run produces a NIfTI label map and JSON evidence covering labels, voxel counts, physical volumes, geometry checks, and runtime details. It is explicitly not for clinical interpretation, diagnosis, treatment planning, or regulatory submission.
Reads a supplied CT or MRI NIfTI volume, consults skill_manifest.yaml for runtime constraints, invokes the upstream monai.bundle run entry point through scripts/run_ctmr.py, and writes outputs under a caller-provided directory. It accepts CT_BODY, MRI_BODY, or MRI_BRAIN, with optional upstream label IDs, inspects the emitted JSON and label map, and records input geometry, output path, observed and unexpected labels, per-class voxel counts, physical volumes derived from header spacing, runtime, upstream command, model inventory, and geometry checks. It can record a ground-truth path but does not compute Dice.
- A medical-imaging engineer needs automatic CT_BODY segmentation with a saved label map and evidence JSON.
- A researcher needs to segment MRI_BODY volumes while retaining geometry and per-class statistics.
- A brain-MRI workflow has completed the upstream preprocessing requirements and needs MRI_BRAIN bundle inference.
- An evaluator needs to record a reference label map and use the paired verifier for optional CT-body Dice or IoU checks.
What are this skill's strengths and limitations?
- Supports automatic CT_BODY, MRI_BODY, and MRI_BRAIN segmentation.
- Produces both a label map and evidence-oriented JSON with geometry, label, and runtime metadata.
- Delegates inference to the upstream pinned MONAI bundle instead of reimplementing it.
- The repository documents CLI installation, daily synchronization, and signature verification.
- Requires an NVIDIA GPU with CUDA and several Python runtime packages.
- Setup may contact GitHub or Hugging Face and use local caches.
- MRI_BRAIN input must already meet the upstream brain-preprocessing requirements.
- It makes no clinical, diagnostic, treatment-planning, or regulatory claims and does not calculate Dice itself.
How do you install this skill?
Install the selected skill from NVIDIA/skills with: npx skills add nvidia/skills --skill nv-segment-ctmr --yes. The README also documents --list and agent-specific installation with --agent codex or --agent claude-code. It does not require manually cloning or copying skill folders.
How do you use this skill?
From the repository root, in an environment with the upstream bundle and model files available, run: export NV_SEGMENT_CTMR_ROOT="${NV_SEGMENT_CTMR_ROOT:-$HOME/.cache/nvidia-skills/upstreams/NV-Segment-CTMR-f9f5f51/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 PATH_TO_IMAGE.nii.gz with the supplied path; use MRI_BODY or MRI_BRAIN as needed. Add --label-prompts "3,14" for selected class IDs or --ground-truth PATH to record a reference label map. Read skill_manifest.yaml first and inspect the output JSON and verifier guidance.