GliomaSAM3-MoE (Minimal BraTS2023 3D Segmentation)

This is a minimal, fully runnable BraTS2023 3D segmentation project with a complete GliomaSAM3_MoE model, synthetic data support, training, inference, and tests.

Install

pip install -r requirements.txt

Run tests

pytest -q

Synthetic debug training (no real data required)

python train.py --config configs/debug.yaml --synthetic true

Real BraTS data layout (expected)

Each case is a folder under data.root_dir, containing:

case_id/
  t1n.nii.gz
  t1c.nii.gz
  t2f.nii.gz
  t2w.nii.gz
  seg.nii.gz

Label values must be in {0, 1, 2, 4}.

SegMamba preprocessed data (npz)

If you use the SegMamba preprocessing pipeline, place *.npz under:

./data/fullres/train

This project supports that format with data.format: "segmamba_npz" (already in configs). It will read *.npz and cached *.npy / *_seg.npy, and automatically map label 3 -> 4.

Recommended paths (aligned with SegMamba):

checkpoints: ./logs/segmamba/model
predictions: ./prediction_results/segmamba

Example:

python train.py --config configs/train.yaml
python infer.py --config configs/train.yaml --input ./data/fullres/train --checkpoint ./logs/segmamba/model/ckpt_stepXXXX.pt --output ./prediction_results/segmamba

Inference

python infer.py --config configs/train.yaml --input /path/to/case_or_root --checkpoint /path/to/ckpt.pt --output ./outputs

Outputs:

*_regions_prob.nii.gz : probability maps for [WT, TC, ET]
*_regions_bin.nii.gz : thresholded binary maps
*_label.nii.gz : final label map in {0,1,2,4}

When data.format: "segmamba_npz", infer.py also writes:

{case_id}.nii.gz : 3-channel (TC/WT/ET) mask for SegMamba 5_compute_metrics.py