GAST-Mamba: Adaptive Gate-Aware Mamba for MRF

GAST-Mamba is the reference implementation behind Adaptive Gate-Aware Mamba Networks for Magnetic Resonance Fingerprinting (Ding et al., 2025). The codebase mirrors the manuscript training recipe: deterministic runs, dataset-wide T1/T2 normalisation, and the asymmetric MSE + L1 loss used for the reported results. A preprint is available at doi:10.48550/arXiv.2507.03369.

Quickstart

conda create -n gast-mamba python=3.11 -y
conda activate gast-mamba
pip install -e .

Launch training with:

scripts/train.sh configs/gast_mamba.yaml

This script sets PYTHONPATH, computes target statistics on-the-fly, and writes checkpoints plus target_stats.json to checkpoints/.

Evaluation & Export

scripts/eval.sh configs/gast_mamba.yaml checkpoints/epoch_0100.pth --split train
python -m gast_mamba.export_nifti checkpoints/epoch_0100.pth src/gast_mamba/data/t200 --extension nifti

Both commands respect ROI masks and reuse the stored normalisation stats, matching the paper’s validation flow.

Data Availability

The curated training dataset contains 192×192×105 NIfTI volumes with real/imaginary fingerprints plus T1, T2, and ROI masks. It is available upon request; contact the authors to obtain access.

Project Layout

configs/          YAML experiment configs
scripts/          Entry points for SVD prep, training, evaluation
src/gast_mamba/   Library code (data loaders, model, losses, utils)

Citation

@article{ding2025gastmamba,
  title   = {Adaptive Gate-Aware Mamba Networks for Magnetic Resonance Fingerprinting},
  author  = {Ding, Tianyi and Contributors, GAST-Mamba},
  journal = {arXiv preprint arXiv:2507.03369},
  year    = {2025}
}

License

Apache-2.0. See LICENSE for full terms.