You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This research release of the model is not a medical device. It is trained on patient-derived data and released under CC BY-NC 4.0 (non-commercial). By requesting access you agree to use it for non-commercial research only and not for clinical decision-making.

ARTIMES: Pleural Mesothelioma Segmentation (nnU-Net ResEnc)

3D volumetric segmentation of pleural mesothelioma on thoracic CT, trained with nnU-Net v2 using the Residual Encoder U-Net (ResEnc) preset.

The model is intended to determine tumor volume in patients already diagnosed with pleural mesothelioma, it is not a diagnostic tool.


Task	Volumetric segmentation of pleural mesothelioma (tumor vs. background) on CT depicting the thorax of patients with the diagnosis pleural mesothelioma
Input	Thoracic CT volume (single channel), `.nii.gz`
Output	Segmentation mask: `0` = background, `1` = Pleural Mesothelioma
Architecture	`ResidualEncoderUNet` (nnU-Net v2), 7 stages, features 32→320
Configuration	`3d_fullres_bs8` (batch size 8) under plans `nnUNetResEncUNetPlans_80G`
Target spacing	3.0 × 0.782 × 0.782 mm (z, y, x)
Patch size	112 × 384 × 384 (z, y, x)
Normalization	CT (HU clipping to foreground percentiles + z-score)
Training data	1,276 CT scans, 5-folds
Framework	nnU-Net v2 (≥ 2.6), PyTorch
License	CC BY-NC 4.0, this research release is not a medical device

Quick start

1. Create a virtual environment

We strongly recommend installing into a fresh virtual environment:

python -m venv artimes-env
source artimes-env/bin/activate        # Windows: artimes-env\Scripts\activate

2. Install PyTorch

Install PyTorch first, following the official instructions for your OS / CUDA version at https://pytorch.org/get-started/locally/, and match the CUDA build to your GPU driver. For example, for a CUDA 12.x driver:

# example, verified on a CUDA 12.x driver:
pip install "torch==2.12.0" --index-url https://download.pytorch.org/whl/cu126

3. Install nnU-Net v2

pip install nnunetv2 huggingface_hub

4. Download these weights

The repo is gated, log in first (hf auth login), then:

hf download nki-radiology/ARTIMES --local-dir ARTIMES

5. Run inference (one line)

Point nnUNet_results at the downloaded folder and run the ensemble of all 5 folds:

nnUNet_results=/../../ARTIMES nnUNetv2_predict -d Dataset804_MPM -i /path/to/input -o /path/to/output -tr nnUNetTrainer -p nnUNetResEncUNetPlans_80G -c 3d_fullres_bs8 -f 0 1 2 3 4

Or use the bundled helper, which sets the environment variable for you:

./predict.sh /path/to/input /path/to/output

Always use the full 5-fold ensemble (-f 0 1 2 3 4), as this is the configuration that was validated.

Input format

nnU-Net requires each case to be a single-channel CT named with the _0000 channel suffix, in the input directory:

input/
  case001_0000.nii.gz
  case002_0000.nii.gz
  ...

Outputs are written as case001.nii.gz, etc., with voxel value 1 marking pleural mesothelioma. A GPU with ≥ 16 GB VRAM is recommended; CPU inference works but is slow. Add --save_probabilities if you need softmax maps.

What's in this repo

Dataset804_MPM/
  nnUNetTrainer__nnUNetResEncUNetPlans_80G__3d_fullres_bs8/
    dataset.json                 # channels + label definitions
    plans.json                   # nnU-Net plans (architecture, spacing, etc.)
    dataset_fingerprint.json     # aggregate training-set statistics (no PHI)
    fold_0/checkpoint_final.pth  # trained weights, fold 0
    fold_1/checkpoint_final.pth
    fold_2/checkpoint_final.pth
    fold_3/checkpoint_final.pth
    fold_4/checkpoint_final.pth
README.md
predict.sh

This is a standard nnU-Net nnUNet_results tree, so the download folder can be used directly as nnUNet_results (or copied into an existing one).

Reproducibility

We verified that these published weights, downloaded fresh from this repo and run with the documented command, reproduce the original training-time validation masks with a voxel agreement of 100.00%.

PyTorch build	CUDA series	Reproduces validation?
2.12.0 + cu126	12.6	✅
2.7.1 + cu128	12.8	✅

Output is identical to the original training-time masks to within a few boundary voxels across these PyTorch versions, so it seems that any correct install reproduces the segmentation.

Guidance:

Install a PyTorch build whose CUDA series your driver supports (nvidia-smi shows the max CUDA version).

Intended use & limitations

This research release of the model is not a medical device and must not be used for clinical decision-making.
Intended to measure tumor volume in patients already diagnosed with pleural mesothelioma from CT depicting the thorax. It is not a diagnostic model and should not be used to detect or diagnose pleural mesothelioma.
Behavior on other pathologies or body regions is undefined.

Citation

If you use these weights, please cite the ARTIMES study and the original nnU-Net paper:

Groot Lipman KBW, Wittenberg R, de Oliveira Taveira M, et al. Development and validation of artificial intelligence-assisted volumetric response criteria in pleural mesothelioma (ARTIMES): a retrospective, multicohort, multicentre study. Lancet Oncol 2026; https://doi.org/10.1016/S1470-2045(26)00084-7

Isensee, F., Jaeger, P.F., Kohl, S.A.A. et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18, 203–211 (2021).

License

Released under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0).

Downloads last month: -; Downloads are not tracked for this model. How to track