NV-Segment-CT Fine-Tuning
Fine-tune NV-Segment-CT VISTA3D on labeled CT NIfTI datasets.
The files disclose permissions, inputs/outputs, network endpoints, cache usage, GPU use, local writes, dependency installation, and the non-clinical boundary, while pinning major upstream revisions. However, Bash access, modification of the active Python environment, writes into the repository bundle, a roughly 832 MB model download, and overseas network access remain broad; per-run confirmation, strong isolation, and rollback are incomplete, so points are deducted.
The script provides MONAI version checks, diagnosable errors for missing dependencies and bundle files, dataset shape/affine validation, and OOM/checkpoint diagnostics. Committed tests cover several key configuration paths, and the benchmark reports validation and runtime results. This review did not execute the skill, and GPU, upstream-bundle, and environment-drift failures remain possible, so the score stays below the static ceiling.
The audience, CT NIfTI inputs, smoke/sanity/user-dataset scenarios, parameters, outputs, and non-fit boundaries are reasonably clear, including the prohibition on inventing new VISTA3D classes. Chinese-language guidance and semantic exclusion triggers are absent, and reachability of the Hugging Face/GitHub-dependent core workflow from mainland China is not addressed, so points are deducted.
SKILL.md is layered with purpose, scripts, prerequisites, usage, data contract, results, limitations, troubleshooting, and verification. The manifest documents versions, dependencies, inputs, outputs, side effects, upstream revisions, and a verifier; Apache-2.0 licensing, versioning, and provenance are mostly clear. The benchmark reports an author-format schema finding, while a formal changelog, named maintenance responsibility, update path, and fully consistent dependency constraints are missing, so points are deducted.
The skill offers a smoke pipeline check, Task06 sanity reproduction, user-data finetuning, data auditing, formal evaluation, checkpoint selection, and structured output.json. The benchmark reports positive-task correctness and effectiveness uplift. Because this review is static and cannot independently confirm output usability, convergence, or benefit over manual alternatives, the score is capped by calibration and remains conservative.
Pinned upstream revisions, reference metrics, script constants, committed tests, a verifier, and the benchmark report provide auditable evidence and distinguish smoke plumbing from quality evaluation. However, execution was not performed here, and the benchmark claims cannot be independently reproduced from this static review, so the score is limited to the static ceiling.
- Execution may install packages into or modify the active Python environment; use an isolated environment explicitly selected by the user and obtain confirmation beforehand.
- Core model and configuration downloads depend on Hugging Face/GitHub; assess mainland-China reachability, mirroring, and supply-chain risk.
- Smoke mode demonstrates plumbing only, not full-dataset convergence; sanity metrics and the non-clinical disclaimer do not replace clinical validation.
- The script writes bundle configuration files and the caller-selected output directory; confirm paths, disk capacity, and recovery procedures before execution.
What does this skill do, and when should you use it?
This skill runs smoke checks or dataset fine-tuning for NV-Segment-CT VISTA3D using CT NIfTI labels. It delegates execution to the upstream MONAI bundle rather than handwritten training or inference code, and it is not intended for clinical validation, deployment, diagnosis, or regulatory submission. It accepts dataset, datalist, anatomy, label-mapping, smoke, sanity, and evaluation options, then produces checkpoints and schema-checked JSON evidence. GPU runs require CUDA-capable PyTorch alongside Python and the documented scientific-computing dependencies.
Runs scripts/run_finetune.py, reading dataset/imagesTr/*.nii.gz, dataset/labelsTr/*.nii.gz, or a datalist and applying dataset_dir, target_anatomy, label_mapping, smoke, sanity, and auto_seg inputs. It stages configuration files, calls the upstream MONAI bundle for training and validation, writes checkpoints and evidence under --output-dir, and creates output.json. It supports smoke workflow checks, MSD Task06 Lung Tumor sanity reproduction, and user-data fine-tuning. Runs may cache Hugging Face assets and contact huggingface.co or raw.githubusercontent.com.
- A researcher who needs to verify wiring, configuration generation, checkpoint loading, and runtime compatibility can run the --smoke workflow on a tiny fixture.
- A medical-AI engineer reproducing the MSD Task06 Lung Tumor single-GPU reference recipe can use --sanity for the five-epoch training and pre/post formal evaluation flow.
- A team with custom CT NIfTI data and label mappings can submit a user dataset through dataset-dir, datalist, target-anatomy, and label-mapping.
- An evaluator auditing whether fine-tuning actually changed or improved the model can inspect formal Dice, training validation traces, the recommended checkpoint, and phase logs.
What are this skill's strengths and limitations?
- Provides explicit smoke, sanity, and user-data fine-tuning entry points.
- Wraps the upstream MONAI bundle and records checkpoints, formal evaluation, training traces, and runtime evidence.
- Pins MONAI 1.4.0 to address a documented compatibility issue with boolean-label fine-tuning loss.
- Includes recommended-checkpoint logic and detection of the epoch-0 checkpoint trap.
- It is for smoke or dataset fine-tuning, not clinical validation or clinical deployment.
- GPU execution requires CUDA-capable PyTorch; out-of-memory cases require smaller patches, cache rates, or worker counts.
- Training, validation, transforms, and checkpointing depend on the upstream bundle; multi-GPU runs additionally require host torchrun support.
- The auto-derived plan is heuristic, and the source provides no broad cross-platform test matrix.
How do you install this skill?
Install the skill from the NVIDIA/skills collection with: npx skills add nvidia/skills --skill nv-segment-ct-finetune --yes. Before running, install the documented dependencies: python -m pip install "monai==1.4.0" "numpy<2" pytorch-ignite einops nibabel scipy typer PyYAML fire huggingface_hub. The README does not document a separate standalone installation flow; SKILL.md requires running from the repository root in place.
How do you use this skill?
From the repository root, run: python skills/nv-segment-ct-finetune/scripts/run_finetune.py PATH_TO_DATASET --smoke --patch-size '[64,64,64]' --output-dir runs/nvseg_smoke. For MSD Task06, run: python skills/nv-segment-ct-finetune/scripts/run_finetune.py /path/to/Task06 --sanity --output-dir runs/nvseg_task06_sanity. For user data, run: python skills/nv-segment-ct-finetune/scripts/run_finetune.py --dataset-dir /path/to/dataset --datalist /path/to/datalist.json --target-anatomy "lung tumor" --auto-seg --epochs 5 --patch-size '[128,128,128]' --output-dir runs/nvseg_user_finetune. Start with output.json; when quality gates matter, run python -m eval_engine.run_trusted skills/nv-segment-ct-finetune --fixture skills/nv-segment-ct-finetune/fixtures/spleen_micro --out runs/nvseg_trusted.
How does this skill compare with similar options?
Compared with handwritten training or inference code, this skill requires the upstream MONAI bundle entrypoint and prohibits replacing it with handwritten implementation during normal use; it does not provide an independent clinical-validation workflow.