PANTHER: FLARE 2026 Task 1 weights (team jehb4ik)
Five-fold plain U-Net ensemble (nnU-Net v2.8.1, 3d_fullres, patch 96x160x160,
spacing 2.0x1.5x1.5 mm, 30.9 M parameters per fold) for pan-cancer lesion
segmentation on CT, trained on the per-source capped Dataset500 subset
(8762 scans) of the FLARE 2026 training data. Used together with a Mahalanobis
mask-shape gate fitted on 172 healthy CT scans.
Paper: PANTHER: PAN-cancer CT Segmentation with Two-tier Healthy-scan Ensemble Rejection (MICCAI 2026 FLARE Challenge, Task 1). Code: https://github.com/Jehb4ik/PANTHER-FLARE2026.
| Metric | Value |
|---|---|
| Lesion DSC, hidden validation | 75.38 % |
| Lesion NSD, hidden validation | 70.09 % |
| False-positive rate on 172 healthy CTs (with gate) | 8.1 % |
| Median / max time per scan (RTX 3090) | 9.2 s / 29.3 s |
| Peak GPU memory (nvidia-smi) | 2614 MB |
Files
plans.json,dataset.json: nnU-Net plan and label definitions.fold_{0..4}/checkpoint_final.pth: the five checkpoints (236 MB each).
Usage
Place the files in docker/model/ of the code repository and build the
container, or load them with nnUNetPredictor.initialize_from_trained_model_folder
(the trainer class nnUNetTrainer_Epoch5000_Lr1e3 from the repository must be
importable).
License
Apache License 2.0. Trained only on data provided by the FLARE 2026 organisers; no external pre-trained weights.