Pre-training and Fine-tuning Transformers for fMRI Prediction Tasks

This repo is the implementation for TFF.

Contents

• Datasets
• Train with our datasets
• HyperParameters
• Cite

Datasets

We currently support the following datasets

• HCP - human connectome project S1200

  • Register at (https://db.humanconnectome.org/)
  • Download: WU-Minn HCP Data - 1200 Subjects -> Subjects with 3T MR session data -> Resting State fMRI 1 Preprocessed
  • Preprocess the data by configuring the folders and run 'data_preprocess_and_load/preprocessing.main()'

• ucla (Consortium for Neuropsychiatric Phenomics LA5c Study)

  • Original version available at (https://openneuro.org/datasets/ds000030/versions/00016)
  • Data after proprocessing will be added soon, for now can download original and preprocess indiependently.

Training

• For gender prediction run 'python main.py --dataset_name S1200 --fine_tune_task binary_classification'
• For age prediction run 'python main.py --dataset_name S1200 --fine_tune_task regression'
• For schezophrenia prediction run 'python main.py --dataset_name ucla --fine_tune_task binary_classification'

Tensorboard support

All metrics are being logged automatically and stored in

TFF/runs

Run tesnroboard --logdir=<path> to see the the logs.

HyperParameters

In the future will be added the exact hyperparameters to reproduce results from the paper.

Citing & Authors

If you find this repository helpful, feel free to cite our publication -

TFF: Pre-training and Fine-tuning Transformers for fMRI Prediction Tasks

@misc{2112.05761,
Author = {Itzik Malkiel and Gony Rosenman and Lior Wolf and Talma Hendler},
Title = {Pre-training and Fine-tuning Transformers for fMRI Prediction Tasks},
Year = {2021},
Eprint = {arXiv:2112.05761},
}

Contact: Gony Rosenman, Itzik Malkiel.