MedNeXt-L — BraTS 2026 Brain Metastasis Segmentation
MedNeXt-L kernel5 checkpoint trained on the BraTS 2026 Brain Metastasis Segmentation training set (fold 0, model_best). Part of a three-model ensemble; see the full pipeline at the GitHub repository.
Model Details
- Architecture: MedNeXt-L with 5×5×5 depthwise convolutions
- Input: 4 MRI modalities — T1n, T1c, T2w, T2f
- Output: 5-class segmentation (background, NETC, SNFH, ET, RC)
- Trainer:
nnUNetTrainerV2_MedNeXt_L_kernel5_BraTS - Checkpoint:
model_best.model+model_best.model.pkl(both files required)
Label Convention
| Label | Region |
|---|---|
| 0 | Background |
| 1 | NETC (Non-Enhancing Tumor Core) |
| 2 | SNFH (Surrounding Non-Enhancing FLAIR Hyperintensity) |
| 3 | ET (Enhancing Tumor) |
| 4 | RC (Resection Cavity) |
Usage
Inference code and full ensemble pipeline: GitHub repository
Install MedNeXt following the instructions at MIC-DKFZ/MedNeXt.
Note: both
model_best.modelandmodel_best.model.pklmust be present in the same folder for inference.
Related Models
Citation
@article{TODO}