briandzt/radiomics_nnUNet

Radiomics nnU-Net for PDAC Detection

This model repository contains the nnU-Net checkpoints for the ISBI 2026 work:

From Global Radiomics to Parametric Maps: A Unified Workflow Fusing Radiomics and Deep Learning for PDAC Detection

Paper page: https://huggingface.co/papers/2602.17986

GitHub code and reproduction instructions: https://github.com/briandzt/dl-pdac-radiomics-global-n-paramaps

Model Description

The workflow combines:

• Stage-1 low-resolution pancreas localization using nnU-Net.
• Voxel radiomics parametric map extraction.
• Global radiomics feature extraction.
• Stage-2 full-resolution PDAC detection using an nnU-Net variant with radiomics channels and global radiomics features.

The public checkpoint layout mirrors the GitHub repository's expected src/nnUNet_results structure so users can download the model files directly into a fresh clone.

Repository Layout

Dataset001_LR/
└── nnUNetTrainer__nnUNetPlans__3d_fullres/
    ├── dataset.json
    ├── dataset_fingerprint.json
    ├── plans.json
    ├── fold_0/checkpoint_best.pth
    ├── fold_1/checkpoint_best.pth
    ├── fold_2/checkpoint_best.pth
    ├── fold_3/checkpoint_best.pth
    └── fold_4/checkpoint_best.pth

Dataset002_stage2/
└── nnUNetTrainer_Loss_CE_checkpoints__nnUNetPlans__3d_fullres/
    ├── dataset.json
    ├── dataset_fingerprint.json
    ├── plans.json
    ├── fold_0/checkpoint_best.pth
    ├── fold_1/checkpoint_best.pth
    ├── fold_2/checkpoint_best.pth
    ├── fold_3/checkpoint_best.pth
    └── fold_4/checkpoint_best.pth

Intended Use

This release is intended for research reproducibility and method comparison for PDAC detection on venous-phase pancreatic CT.

The workflow outputs:

• a voxel-level PDAC detection map
• a case-level PDAC likelihood JSON file

This model is not intended for clinical deployment or standalone diagnosis.

How to Use

Clone the GitHub code repository, then download the checkpoints:

powershell -ExecutionPolicy Bypass -File .\scripts\download_checkpoints_from_hf.ps1 `
  -RepoId briandzt/radiomics_nnUNet

Verify the local setup without a GPU:

python scripts/verify_repository.py

Run local inference:

python main.py -i ./workspace/test_example/input -o ./workspace/test_example/output --image-ext .nii.gz

Run Docker inference:

docker build -t pdac-radiomics-paramaps .
docker run --gpus all --rm \
  -v /path/to/input:/input:ro \
  -v /path/to/output:/output \
  pdac-radiomics-paramaps

Inputs and Outputs

Expected input:

• venous-phase pancreatic CT
• NIfTI (.nii.gz) for local runs or MHA (.mha) for challenge-style Docker input

Expected output:

output/
├── <case>_pdac-likelihood.json
└── images/
    └── pdac-detection-map/
        └── <case>_detection_map.mha

Training Data and Privacy

This model repository does not include patient imaging data. It contains trained model weights and nnU-Net metadata only.

Limitations

• Requires the GitHub code repository for preprocessing, radiomics extraction, inference, and postprocessing.
• Full inference requires GPU-enabled PyTorch.
• Performance depends on image acquisition, preprocessing compatibility, and domain match to the development data.

Citation

Please cite the paper associated with https://huggingface.co/papers/2602.17986.

@article{deng2026globalradiomicsparametricmaps,
  title = {From Global Radiomics to Parametric Maps: A Unified Workflow Fusing Radiomics and Deep Learning for PDAC Detection},
  year = {2026},
  note = {Hugging Face Papers: https://huggingface.co/papers/2602.17986}
}

License

License details should be finalized before public release. If the GitHub repository uses a specific license, mirror that license here.