Evaluation of Task-fMRI Decoding with Deep Learning on a Small Sample Sized Dataset

This is a repository of codes for Evaluation of Task-fMRI Decoding with Deep Learning on a Small Sample Sized Dataset.

requirements

• scikit-learn == 0.22.2.post1
• pytorch == 1.3.1
• captum == 0.2.0
• nipy == 0.4.2
• nilearn == 0.6.2

1. Prapare dataset

We should set config.ini for data directory, fMRI image file name, label file name, and subjects id.

python scripts/prepare_dataset.py

2. Train classifiers

We can select three split methods and four train methods.

python scripts/train.py [split_method] [train_method]

args	options
split method	LeaveOneSubjectOut, SessionShuffleSplit, SampleShuffleSplit
train method	PLR, SVM, M2DCNN, 3DCNN

added other split methods: LeaveOneSubjectOut_SessionShuffleSplit, LeaveOneSubjectOut_SampleShuffleSplit

3. Analyze results

3.1 Permutation test

We can calculate the chance level for each split methods.

python scripts/permutation.py [split_method] [permute_number]

3.2 Calculate Integrated Gradients

We can calculate Integrated Gradients for DL model.

python scripts/BackProp.py

3.3 Aggregate and Visualize results

python scripts/visualize.py
python scripts/aggregate.py

3.4 (post hoc) Clustering and T-SNE

python scripts/clustering.py