Here, we provide a set of codes to build and train a weight-sparsity-controlled deep neural network (DNN) model to identify individuals in the HCP 1200 dataset using time-varying functional connectivity (tvFC) patterns during a resting state.
The codes were implemented in the Python 3.6 environment with the following libraries:
- TensorFlow 1.15
- numpy, timeit, zipfile, os, datetime, pytz, scipy, functools, matplotlib
This main code is for identifying 10 example subjects using 15-s window tvFC patterns.
Hidden layers are initialized using a pretrained model to shorten a training time.
This module contains the classes/functions used in the main code.
This code is a function to transform 1D vectors of FC (input tvFC or fpFC) to 2D matrices.
This directory includes the input tvFC data.
The subdirectories RS1 (or Day1) and RS2 (or Day2) indicate resting-state fMRI (rfMRI) sessions from the first and second visits, respectively. In each visit, two rfMRI runs were acquired with different phase encoding directions, i.e., right-to-left (RL) and left-to-right (LR).
This directory contains example results of 'main_indiv_identification.ipynb'.
Similarly, the main code will store the corresponding results in an automatically created folder under the './results' directory.
* Please note that 'result_weight.mat' was not included here because it was too large to upload to gitHub.
This directory contains example fpFCs obtained from the trained model of example results.
Juhyeon Lee, Ph.D. candidate
jh0104lee@gmail.com
Brain Signal Processing Lab
Department of Brain and Cognitive Engineering, Korea University, Seoul, Republic of Korea