nv-segment-ctmr

nv-segment-ctmr is a thin wrapper skill from the NVIDIA MedTech Team that runs the upstream NV-Segment-CTMR (VISTA3D) MONAI segmentation bundle on CT or MRI NIfTI volumes and records the resulting label-map as evidence. It solves the problem of reproducibly invoking a complex medical-imaging model: rather than reimplementing inference, the skill shells out to the documented `python -m monai.bundle run` entrypoint and then inspects the produced NIfTI label map against an output schema. The documentation is emphatic that this is not for clinical interpretation, diagnosis, regulatory submission, or treatment planning.

The workflow centers on a single entrypoint, `scripts/run_ctmr.py`, which takes an input volume, a modality flag (CT_BODY, MRI_BODY, or MRI_BRAIN), and an output directory, and can optionally accept label prompts. It performs automatic "segment everything" inference, loads label names from upstream configs when available, and emits a JSON result plus verifier guidance. A skill_manifest.yaml declares inputs (ct_or_mr_volume), outputs (label_map, result_json), validation gates, and side effects such as caching model assets under the Hugging Face cache and contacting github.com or huggingface.co during one-time setup. GPU/CUDA may be required, and a pinned set of Python packages (monai, numpy<2, nibabel, scipy, etc.) must be installed.

The target users are medical-imaging engineers, MONAI/VISTA3D practitioners, and ML pipeline builders who need a deterministic, evidence-recording wrapper around CT/MRI whole-body or brain segmentation. It is explicitly a research/engineering harness, not a clinical tool, so its realistic use cases are benchmarking, pipeline integration, and label-map evidence generation rather than any patient-facing decision.