MICCAI BraTS-METS 2026 — Task 1 checkpoints
Fine-tuned nnU-Net v2 checkpoints for brain metastases segmentation.
Code, inference pipeline and full method description:
https://github.com/thekeramik05/MICCAI-METS_challenge_bi-intelContents
weights.zip contains 7 fold checkpoints only:
| Model | Trainer | Folds |
|---|
| A ("Triad") | nnUNetTrainerBraTS_TriadInit_SmallLesionWeightedCE_CEw3__3_FT900 | 0–4 |
| B ("BrainIAC") | nnUNetTrainerBraTS_BrainIACWrapper_RCOversample__3 | 1, 3 |
plus the matching dataset.json / plans.json.
The Triad and BrainIAC foundation checkpoints are not redistributed here.
Obtain them from their authors and set TRIAD_CKPT / BRAINIAC_CKPT — see §3.1
of the GitHub README. BRAINIAC_CKPT is required at inference time as well as
during training.
What Model B actually contains
Model B checkpoints are not standalone nnU-Net weights. Each embeds a complete,
frozen copy of the BrainIAC ViT-B/16 backbone:
| Module prefix | Parameters | Origin |
|---|
base_network.* | 88.2 M | nnU-Net, trained here |
brainiac.* | 116.7 M | BrainIAC ViT-B/16, verbatim |
proj.* | 3 K | projection layer, trained here |
| total | 204.9 M | |
Model A contains no third-party weights: the Triad checkpoint initialised the
encoder, which was then fully fine-tuned, leaving a plain 88.2 M nnU-Net.
Provenance
License
| Component | Terms |
|---|
| Model A checkpoints | Derived from Triad (MIT) — no additional restriction |
| Model B checkpoints | BrainIAC Research-Only License (see LICENSE) |
Because Model B embeds the BrainIAC backbone, use of these weights is permitted
solely for non-commercial academic research and educational purposes.
Commercial use — including clinical workflows, decision-support systems,
healthcare operations, or any fee-bearing product or service — is prohibited
without a separate written licence from Mass General Brigham, which reserves all
commercial rights.
