Used for running NV-Segment-CT VISTA3D on CT NIfTI volumes and recording label-map evidence.
NV-Segment-CT is an NVIDIA MedTech skill for running NV-Segment-CT VISTA3D segmentation on CT volumes in NIfTI format and recording label-map evidence for downstream verification workflows. It is explicitly not for clinical interpretation, autonomous diagnosis, or regulatory submission. It solves the problem of reliably invoking a complex upstream medical-imaging inference pipeline by providing a thin, documented wrapper with strong validation gates instead of a handwritten reimplementation.
The single entrypoint, run_vista3d.py, takes a CT NIfTI path plus optional output directory and label-prompt IDs, and delegates all inference, preprocessing, and postprocessing to the official Hugging Face pipeline helper bundled under bundle/. It creates an isolated Python 3.10 virtual environment rather than installing model dependencies into the caller's active environment, auto-detects CUDA versus CPU (overridable with a device flag), and pins compatibility versions such as transformers 4.46.3 against the upstream Torch 2.0.1. Outputs are a NIfTI label-map and a structured JSON evidence record with per-class voxel counts, physical volumes, geometry validation, and runtime metadata. Sanity gates assert at least one foreground voxel per requested anatomy so that schema-valid but semantically empty output is caught. The skill ships a manifest, output schema validator, tests, fixtures (including an optional spleen fixture fetcher), and an NVSkills-Eval benchmark that scored 100% on the security dimension.
It targets developers and engineers building or verifying CT-segmentation workflows in research and development. First-run setup may contact Hugging Face and raw.githubusercontent.com to download the model bundle, and the optional spleen fixture is fetched from a public S3 bucket; the docs warn against loading untrusted checkpoints into the legacy reproduction environment.
Input is a CT volume in NIfTI (.nii.gz) format, optionally with specific label-prompt IDs. Output is a NIfTI label-map plus a structured JSON evidence record containing per-class voxel counts, physical volumes, geometry validation, and runtime metadata.
No. It is explicitly documented as not for clinical deployment, clinical interpretation, autonomous diagnosis, or regulatory submission. It is intended to produce label-map evidence for research and development verification workflows.
The skill creates a dedicated Python 3.10 virtual environment under the user's cache directory and invokes its binaries directly, so it does not install model dependencies into the caller's active environment. It pins compatibility versions against the upstream Torch 2.0.1.
First setup may contact Hugging Face and raw.githubusercontent.com to download the model bundle and cache assets. The optional spleen fixture fetcher downloads the MSD09 dataset from a public AWS S3 bucket. The docs warn against loading untrusted checkpoints into the legacy environment.
Output can be schema-valid but semantically empty, for example when label prompts do not match the input anatomy. Sanity gates assert at least one foreground voxel per requested anatomy, and troubleshooting guidance advises re-running with a known fixture and inspecting the JSON and stderr.
Quick Setup:
.claude/skills/Repository
nvidia/skills