NV-Segment-CT Finetune

Purpose

• Used for smoke or dataset finetuning of NV-Segment-CT VISTA3D on CT NIfTI labels, including the upstream fixed-channel softmax workflow and optional MLflow tracking. Not for clinical validation.
• Wraps the upstream MONAI bundle entrypoint; do not replace it with handwritten training or inference code.
• Manifest inputs are datasetdir, datalist, targetanatomy, labelmapping, smoke, sanity, autoseg, softmax, skipformaleval, mlflowtrackinguri, mlflowexperimentname, and mlflowrunname.
• Manifest outputs are finetunedckpt and schema-checked resultjson.

Instructions

• Run scripts/run_finetune.py; do not patch files under bundle/ or upstream checkouts during normal skill use.
• For standalone Bash, include the fresh-environment setup line before the wrapper; benchmark venvs start empty.
• Run the committed script in place from the repo root. Do not copy this skill to a runtime directory, and do not use rm or cleanup commands in generated invocations.
• If a host exposes runscript, use runscript("scripts/run_finetune.py", args=[...]); otherwise run from the repo root.
• For the shortest workflow check, use --smoke; for MSD Task06 Lung Tumor reproduction, use --sanity.
• Choose between the standard and --softmax workflows using the criteria below. Do not combine --softmax with --auto-seg or --sanity.
• Set --mlflow-experiment-name to enable MLflow for the training phase of either workflow. --mlflow-tracking-uri and --mlflow-run-name require an experiment name. Formal pre/post evaluation does not receive MLflow credentials.
• Read references/task06-and-results.md only when you need Task06 reference details, output-field definitions, or manual bundle setup notes.

Choosing the Workflow

Use --softmax only when all of these conditions hold:

• The complete class set is known before training and will not vary between inference requests.
• Labels are mutually exclusive: each voxel is background or exactly one foreground class.
• Every foreground dataset label maps to an existing VISTA3D class ID, and a conventional fixed-channel output is desired.

Keep the standard workflow if point prompts must remain available, classes are selected dynamically at inference, labels can overlap, or the Task06 --sanity reproduction is required.

For --label-mapping '[[1,3],[2,13]]', channel 0 is background, channel 1 represents dataset label 1 initialized from VISTA3D class 3, and channel 2 represents dataset label 2 initialized from VISTA3D class 13. Preserve the entries and their order when using the resulting modelsoftmax.pt with upstream configs/inferencesoftmax.json. The nv-segment-ct and nv-segment-ctmr inference skills do not currently expose that fixed-channel inference path.

Available Scripts

Prerequisites

• Python 3.10+ with CUDA-capable Torch for GPU runs.
• Runtime packages from skillmanifest.yaml, especially monai==1.4.0, numpy<2, nibabel, scipy, typer, PyYAML, fire, pytorch-ignite, einops, and huggingfacehub. Install mlflow>=2.10,<4 when MLflow tracking is enabled.
• Optional environment variables: CUDAVISIBLEDEVICES restricts visible GPUs; NPROCPERNODE overrides GPU count and values >=2 select multi-GPU mode for non-sanity runs; NVSEGFINETUNEAUTOVENV=0 disables the cached MONAI 1.4 compatibility environment. Remote tracking may use DATABRICKSCONFIGPROFILE, DATABRICKSHOST, DATABRICKSTOKEN, MLFLOWTRACKINGCLIENTCERTPATH, MLFLOWTRACKINGINSECURETLS, MLFLOWTRACKINGPASSWORD, MLFLOWTRACKINGSERVERCERTPATH, MLFLOWTRACKINGTOKEN, or MLFLOWTRACKINGUSERNAME; these variables are forwarded only when MLflow is explicitly enabled, and unrelated credentials are not forwarded.
• --softmax also needs the pinned NVIDIA-Medtech source checkout. Set NVSEGMENTCTROOT to its NV-Segment-CT directory, or set NVSEGMENTCTMRROOT to the sibling NV-Segment-CTMR directory. The wrapper reads the official softmax config and implementation in place and writes generated overrides only under --output-dir.
• Side effects: writes generated bundle configs under skills/nv-segment-ct-finetune/bundle/configs/, including skills/nv-segment-ct-finetune/bundle/configs/autooverride.json, skills/nv-segment-ct-finetune/bundle/configs/traincontinualtask06lung.json, and skills/nv-segment-ct-finetune/bundle/configs/dfwnologging.json; writes checkpoints/evidence under --output-dir and local tracking data under <output-dir>/mlruns when enabled; may create the MONAI compatibility environment under ~/.cache/nvidia-skills/venvs/nv-segment-ct-finetune-monai14/; may cache model assets under ~/.cache/huggingface/; and may contact https://huggingface.co, https://raw.githubusercontent.com, or https://<caller-provided-mlflow-or-databricks-workspace> when remote tracking is explicitly enabled.

Fresh environment setup:

python -m pip install "monai==1.4.0" "numpy<2" pytorch-ignite einops nibabel scipy typer PyYAML fire huggingface_hub

When MLflow tracking is enabled, also install:

python -m pip install "mlflow>=2.10,<4"

Known upstream compatibility constraints:

• DFW Task06 reference: Python 3.10.16, MONAI 1.4.0, Torch 2.7.0+cu126.
• Use exact monai==1.4.0 for smoke, sanity, and evidence runs; MONAI 1.5.x can crash the upstream finetune loss on boolean labels.
• Do not float the dependency as monai>=1.4,<1.6 in generated commands.
• The softmax workflow keeps the upstream defaults of 100 epochs and learning rate 1e-4 unless the caller overrides them.

One-time source setup for --softmax:

export NV_SEGMENT_CTMR_COMMIT=cb921f5c58837c0f42a713855d68b32af88e1cdd
export NV_SEGMENT_CTMR_CHECKOUT="$HOME/.cache/nvidia-skills/upstreams/NV-Segment-CTMR-cb921f5"
if [ ! -d "$NV_SEGMENT_CTMR_CHECKOUT/.git" ]; then
  git clone https://github.com/NVIDIA-Medtech/NV-Segment-CTMR.git "$NV_SEGMENT_CTMR_CHECKOUT"
fi
git -C "$NV_SEGMENT_CTMR_CHECKOUT" checkout --detach "$NV_SEGMENT_CTMR_COMMIT"
export NV_SEGMENT_CT_ROOT="$NV_SEGMENT_CTMR_CHECKOUT/NV-Segment-CT"

Usage

Smoke-scale workflow check:

python -m pip install "monai==1.4.0" "numpy<2" pytorch-ignite einops nibabel scipy typer PyYAML fire huggingface_hub && \
python skills/nv-segment-ct-finetune/scripts/run_finetune.py \
  PATH_TO_DATASET \
  --smoke \
  --patch-size '[64,64,64]' \
  --output-dir runs/nvseg_smoke

Use the staged dataset as PATHTODATASET. For the micro fixture, use skills/nv-segment-ct-finetune/fixtures/spleen_micro. Smoke mode proves wiring, config generation, checkpoint loading, and runtime compatibility; it is not a quality bar.

MSD Task06 Lung Tumor sanity reproduction:

python skills/nv-segment-ct-finetune/scripts/run_finetune.py \
  /path/to/Task06 \
  --sanity \
  --output-dir runs/nvseg_task06_sanity

The sanity preset follows the single-GPU DFW recipe: fold-0 validation, label mapping [[1, 23]] for lung tumor, automatic class-prompt segmentation, patch [128,128,128], 5 epochs, and original-spacing configs/evaluate.json scoring before and after training. Expected reference range is pretrained Dice about 0.6697, training-best Dice about 0.6905, and fine-tuned formal Dice about 0.6836.

User-data finetune:

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

Use --label-mapping '[[1, 23]]' when local label values are custom or the anatomy name is ambiguous.

Optional local MLflow tracking:

python skills/nv-segment-ct-finetune/scripts/run_finetune.py \
  --dataset-dir /path/to/dataset \
  --datalist /path/to/datalist.json \
  --target-anatomy "lung tumor" \
  --epochs 5 \
  --mlflow-experiment-name nvseg-finetune \
  --mlflow-run-name trial-01 \
  --output-dir runs/nvseg_mlflow

This uses MONAI's documented --tracking mlflow path and built-in rank-zero handlers. With no --mlflow-tracking-uri, data stays in <output-dir>/mlruns. Pass a caller-approved remote URI, including databricks, only when remote tracking is intended. MLflow does not change patch size, transforms, optimizer values, DataLoader settings, or other training configuration.

Fixed-channel softmax finetune for mutually exclusive labels:

export NV_SEGMENT_CT_ROOT="$HOME/.cache/nvidia-skills/upstreams/NV-Segment-CTMR-cb921f5/NV-Segment-CT"
python skills/nv-segment-ct-finetune/scripts/run_finetune.py \
  --dataset-dir /path/to/dataset \
  --datalist /path/to/datalist.json \
  --label-mapping '[[1,3],[2,13]]' \
  --softmax \
  --epochs 100 \
  --output-dir runs/nvseg_softmax

This delegates to upstream configs/traincontinualsoftmax.json. It produces checkpoints/modelsoftmax.pt; the source model.pt initializes the network but is not compatible with configs/inferencesoftmax.json. The wrapper therefore recommends the produced softmax checkpoint after a successful run.

Examples

Smoke run on a staged tiny dataset:

python skills/nv-segment-ct-finetune/scripts/run_finetune.py \
  runs/with_vs_without_nv/_inputs/nv_segment_ct_finetune/input_dataset \
  --smoke \
  --patch-size '[64,64,64]' \
  --output-dir runs/nvseg_smoke

Task06 sanity run on a local MSD cache:

python skills/nv-segment-ct-finetune/scripts/run_finetune.py \
  .workbench_data/datasets/Task06_Lung \
  --sanity \
  --output-dir runs/nvseg_task06_sanity

Data Contract

• Preferred layout: dataset/imagesTr/.nii.gz and dataset/labelsTr/.nii.gz.
• Labels must align one-to-one with images by basename.
• The target label value must be present in the training labels.
• Use a datalist when patient-level splitting matters. The bundle default fold is 0, so fold: 0 entries are validation and all other folds are training.
• Every trained foreground label must map to an existing VISTA3D global class id from bundle/label_dict.json; this skill cannot invent a new class.
• In --softmax mode, the first mapping column is the saved dataset label and the second is the pretrained VISTA class ID. Mapping order fixes the channel layout and must remain unchanged during inference.

Results

Check output.json in the run directory first:

• formalpretrainedvaldice and formalfinetunedvaldice: original-spacing pre/post scores when formal eval is enabled.
• trainingstartvaldice, valdiceperepoch, and trainingbestval_dice: training-time validation trace.
• finetunedckptmatchespretrainedweights: detects the standard workflow's epoch-0 checkpoint trap when valatstart=true; softmax uses a different checkpoint architecture.
• recommendedckpt: checkpoint to keep. Do not blindly use the last epoch, modelfinetune.pt, or model_softmax.pt without checking the recorded workflow and metrics.
• invocation.mlflow_tracking: selected tracking URI, experiment name, and optional run name, or null when tracking was disabled.
• runtime.oom, runtime.peakgpumb, and phase logs: distinguish OOM, slow validation, and process failure.

Decision rule: prefer formal original-spacing pre/post scores when present; reject tensor-identical "fine-tuned" checkpoints for sanity recovery; treat improved: false as valid evidence rather than a wrapper failure.

Limitations

• Thin wrapper. Training, validation, transforms, and checkpointing are delegated to the upstream bundle in bundle/.
• Reproduction record only: the successful five-epoch Task06 run used Python

3.12.3, PyTorch 2.12.0+cu130 with CUDA 13.0, MONAI 1.4.0, NumPy 1.26.4, PyTorch-Ignite 0.5.4, NiBabel 5.4.2, SciPy 1.16.0, einops 0.8.2, Fire 0.7.1, Hugging Face Hub 0.36.2, Transformers 4.57.6, Typer 0.25.1, PyYAML 6.0.3, and MLflow 3.14.0 on one NVIDIA RTX 6000 Ada 48 GB GPU. These versions document the evidence environment; they are not additional package constraints or a claim that other versions cannot work.

• The auto-derived plan is heuristic; caller-provided --patch-size, --cache-rate, --epochs, and --learning-rate win.
• --softmax is not compatible with --sanity: the Task06 reference scores and original-spacing pre/post evaluation belong to the standard VISTA3D continual-learning workflow. Softmax runs record the training validation trajectory but need a separate task-specific evaluation before quality claims.
• The Task06 sanity recipe intentionally forces single-GPU execution to match the DFW reference. Multi-GPU mode for other datasets requires host torchrun support.
• The paired verifier is CPU-only and audits the evidence pack; it does not re-run GPU segmentation.
• MLflow support is optional and uses MONAI's built-in tracking handlers. Tracking errors are part of the upstream MONAI run and can therefore fail the finetune command.
• Not for clinical deployment, clinical interpretation, autonomous diagnosis, or regulatory submission.

Troubleshooting

Verification

Run the implemented verifier when quality gates matter:

python -m eval_engine.run_trusted skills/nv-segment-ct-finetune \
  --fixture skills/nv-segment-ct-finetune/fixtures/spleen_micro \
  --out runs/nvseg_trusted