DyBraSS

Official PyTorch implementation of DyBraSS: A Unified Spatiotemporal State-Space Model for Dynamic Brain State Analysis in Resting-State fMRI.

Installation

conda create -n dybrass python=3.10
conda activate dybrass
pip install -r requirements.txt

Repository structure

conf/            Hydra configurations
dataset/         Data loading utilities
models/dybrass.py  DyBraSS implementation
training/        Training and evaluation code
utils/           Learning-rate scheduler
scripts/         Preprocessing and launch scripts
main.py          Training and evaluation entry point

Data preparation

Prepare a NumPy archive containing:

• dfc: an object array of length N; each element has shape [L_i, R, R]
• label: an integer array of shape [N]

To generate dFC sequences from ROI time series:

python scripts/prepare_dfc.py \
  --input /path/to/roi_timeseries.npz \
  --output /path/to/dfc.npz

Split files use archive row indices. To generate a new five-fold split:

python scripts/make_splits.py \
  --data /path/to/dfc.npz \
  --output splits/custom/seed_42.json \
  --seed 42

Training

python main.py \
  experiment=abide \
  dataset.data_path=/path/to/abide_dfc.npz \
  dataset.split_path=/path/to/seed_42.json

Use experiment=adhd or experiment=cobre for the other datasets. To run all predefined seeds:

bash scripts/run_all_seeds.sh \
  abide \
  /path/to/abide_dfc.npz \
  /path/to/abide_splits