NV-Segment-CT

132-Class CT Segmentation

Commercial

NV-Segment-CT (VISTA3D) is an interactive foundation model for 3D CT segmentation covering 132 anatomical structures. It offers three workflows: segment everything for whole-body exploration, segment by class for specific structures, and point-click prompts to refine results and speed up ground-truth annotation. It takes a CT volume in NIfTI format and returns a 3D label map.

CT HuggingFace GitHub Paper

Agent skills

Run it with an agent

Install the skill, then just ask your agent in plain language. medagent.nv_segment_ct

install skill

npx skills add NVIDIA-Medtech/medical-AI-skills --skill nv-segment-ct

Try saying

“Segment spleen, liver, right kidney, and left kidney from my CT NIfTI volume at /data/case.nii.gz using NV-Segment-CT.”

“Run NV-Segment-CT on data/abdomen.nii.gz and give me the full label map with the structures it found.”

“Use NV-Segment-CT to segment the lungs and heart from my chest CT NIfTI and write the label map to runs/case/.”

Skill on GitHub → Validation: Benchmark·Evals·Tests·Validators How agent skills work

Finetune it with an agent

Install the skill, then just ask your agent in plain language. medagent.nv_segment_ct_finetune

install skill

npx skills add NVIDIA-Medtech/medical-AI-skills --skill nv-segment-ct-finetune

Try saying

“Prepare the shortest smoke-scale NV-Segment-CT finetune command for my MSD-style CT segmentation dataset at /data/msd_case, with outputs under runs/nvseg_msd_case_smoke. This is a command-shape review only; do not install packages or start training in this environment.”

“Prepare a reproducible NV-Segment-CT fixed-channel softmax finetune command for my mutually exclusive two-label CT dataset. The dataset root is /data/two_label_ct, the datalist is /data/two_label_ct/dataset.json, dataset label 1 maps to VISTA3D class 3, dataset label 2 maps to class 13 in that order, and outputs should go under runs/nvseg_softmax. This is a command-shape review only; do not install packages, clone repositories, or start training in this environment.”

“Prepare an NV-Segment-CT finetune command for my CT dataset at /data/lung_ct with datalist /data/lung_ct/dataset.json. Track the run locally in MLflow experiment nvseg-review with run name trial-01, and write outputs under runs/nvseg_mlflow. This is a command-shape review only; do not install packages or start training in this environment.”

“Finetune NV-Segment-CT VISTA3D on my labeled CT dataset at /data/liver_ct using the single-GPU smoke recipe and report the before/after Dice.”

Skill on GitHub → Validation: Benchmark·Evals·Tests·Validators How agent skills work

Engineering & research use only — not a medical device, not for diagnosis or clinical decision-making. Skills follow the open Agent Skills spec.