nv-segment-ct-finetune

This NVIDIA MedTech skill wraps smoke-scale and dataset finetuning of the NV-Segment-CT VISTA3D segmentation model on CT NIfTI labels. It solves the problem of driving an upstream MONAI bundle training entrypoint reproducibly: the wrapper stages configs, runs MONAI, and writes a schema-checked result JSON, and the skill instructs the agent not to replace it with handwritten training or inference code or patch upstream bundle files during normal use. Manifest inputs include the dataset directory, datalist, target anatomy, label mapping, and mode flags; outputs are a finetuned checkpoint and a validated result JSON.

The primary entrypoint, `run_finetune.py`, supports a `--smoke` mode for a fast workflow check and a `--sanity` mode that reproduces the MSD Task06 Lung Tumor recipe (fold-0 validation, label mapping from local class to VISTA3D class 23, automatic class-prompt segmentation, a documented patch size, learning rate, and epoch count, with before/after Dice scoring). Prerequisites are precise: Python 3.10+ with CUDA Torch, pinned `monai==1.4.0` and `numpy<2` plus several supporting packages, with a warning that MONAI 1.5.x can crash the upstream loss on boolean labels. Documented side effects are writing generated bundle configs and checkpoints, caching model assets under the Hugging Face cache, and contacting huggingface.co or raw.githubusercontent.com; a reference file shows manual bundle download via `hf download` and a pinned `urllib` fetch of a label dictionary. The included skill card marks it research-and-development only, reports evaluation results, and lists no required credentials.

Target users are medical-imaging engineers finetuning VISTA3D on custom CT segmentation datasets. The skill states it is not for clinical validation.