This the official Pytorch implementation of the paper fMRI-S4: Learning short- and long-range dynamic fMRI dependencies using 1D Convolutions and State Space Models
This repository requires Python 3.8+ ,Pytorch 1.9+ and Pytorch Lightning
Other packages are listed in requirements.txt.
The Abide dataset can be requested from here The Mddrest dataset can be requested from here
The Mddrest provide the pre-processed timecourses directly. For ABIDE you can generate the time-courses using nilearn.
You can see an example of how to parcellate the data and extract the timecourses in 'data_notebook/parcellate.ipynb'
To integrate it directly to the code, you can organize the paths and into a csv file as in provided in the examples in 'csvfiles/'
All logic for creating and loading datasets to the model is in datasets.py.
The S4 module relies on the "Cauchy kernel" described in the paper.
The implementation of this requires a custom CUDA kernel.
To install it, Run python setup.py install from the directory extensions/cauchy/.
Examples of how to train the model using different configurations
python main.py --dataset ABIDE --atlas HO --n_conv_layers 1 --n_s4_layers 2
python main.py --dataset Mddrest --atlas AAL --n_conv_layers 3 --n_s4_layers 0
The default parameters are described in the paper. You can play around to find the best configuration that works best for your dataset. I observed a relatively robust performance using different hyper-parameters configurations. I recommended using the weights&biases sweep function for hyper-parameters search.
Not written in the paper. I have experimented with visualizing the salient ROIs using different methods in the Captum Library. Feature Permutation with a mask covering the entire temporal profile of regions seems to work best for me. However this needs more exploration. I have tried with simple experiments where I create synthetic classes by classify fMRI with corrupted vs uncorrupted time-courses in certain regions and it seems to work fine. See 'interpret.py' for an example.
A large body of the code used for the S4 model was adopted from this repo, so please if you find any part of this work valuable cite their work along with ours.
@misc{https://doi.org/10.48550/arxiv.2208.04166,
title = {fMRI-S4: learning short- and long-range dynamic fMRI dependencies using 1D Convolutions and State Space Models},
doi = {10.48550/ARXIV.2208.04166},
author = {El-Gazzar, Ahmed and Thomas, Rajat Mani and Van Wingen, Guido},
publisher = {arXiv},
year = {2022},
}
@inproceedings{gu2022efficiently,
title={Efficiently Modeling Long Sequences with Structured State Spaces},
author={Gu, Albert and Goel, Karan and R\'e, Christopher},
booktitle={The International Conference on Learning Representations ({ICLR})},
year={2022}
}