OmniLesion: trained weights
Weights of the OmniLesion submission to the MICCAI FLARE 2026 pan-cancer segmentation task (Task 1). Code, training recipe and evaluation scripts: https://github.com/Nucleo-Research/omnilesion. On the challenge's hidden validation set (100 cases) the pipeline scores 0.8122 lesion DSC and 0.7642 lesion NSD.
Files
| path | content |
|---|---|
model/plans.json, model/dataset.json | nnU-Net v2 plan (residual-encoder U-Net, 3 x 2 x 2 mm, 96 x 128 x 128 patches) and dataset description |
model/fold_0/checkpoint_final.pth | main network after 4000 epochs x 750 iterations (3.0 M updates), with optimiser state; trainer nnUNetTrainerOmniLesion |
lung/detector.pt | RetinaNet lung-nodule detector (MONAI, 3D ResNet-50 + FPN) after 300,000 iterations |
lung/segmentor_state_dict.pt, lung/segmentor_manifest.json | 65^3 nodule patch U-Net (12 base channels) after 40,000 iterations |
sha256.txt lists the checksums. All networks were trained from scratch on the challenge's labelled training data
only (no pretrained weights, no external data, no pseudo labels). These are exactly the weights inside the submitted
container; only the trainer and plan names stored in the checkpoint were renamed to the ones used by the code
repository.
Use
git clone https://github.com/Nucleo-Research/omnilesion.git && cd omnilesion
pip install -r requirements-inference.txt && python -m omnilesion.nnunet_trainer.install
hf download Nucleo-Research/omnilesion --local-dir weights # this repository
python scripts/08_predict.py --inputs /data/images --outputs /data/masks \
--model weights/model --lung-dir weights/lung
Inputs are CT volumes <case>_0000.nii.gz; outputs are binary lesion masks <case>.nii.gz. The pipeline (sliding
window with mirroring, confidence and empty rules, lung-nodule cascade behind a 3.5 L lung gate) and every threshold
are described in the code repository's README and in the paper Pan-cancer Lesion Segmentation in CT under
Heterogeneous and Partial Annotation (FLARE 2026 proceedings).
License
Weights: CC BY-NC 4.0, non-commercial research use, access on request (see the form above). Code: Apache-2.0 at the GitHub repository. Built with nnU-Net v2 and MONAI (both Apache-2.0).