/EEG-classification

Primary LanguagePython

Classifying Functional Neuroimaging Signals

The Dual Stream Neural Network (DSNN) is an end-to-end subject-independent neural network architecture designed for classifying multi-channel functional neuroimaging signals. Some of its salient features include:

  • End-to-end prediction without spending excessive labor on handcrafting/selecting features
  • Automatic feature extraction via deep layer representations in deep layers
  • Very fast predictive capabilities on trained networks

Specific details on the architecture, implementation and results of the DSNN, including those applied to publicly available MEG (data sets 3), EEG, and fNIRS datasets are discussed in detail in the Authors' paper: Dual Stream Neural Networks for Brain Signal Classification

Code Overview

  • preprocess.py: Functions for preprocessing data.
  • visual.py: Some functions for visualization purpose.
  • loss.py: Custom losses.
  • modules.py: Custom network layers that will be called in architectures.py
  • architectures.py: Architectures used in training.
  • Utils.py: Some Utilility functions for different training style.
  • train.py: Training script contains flags for different types of training including:
    * 2D flag: Treating the multi-channel eegs as 2d images;
    *aug flag: Training with generators. Randomly roll the signal or randomly select reconstructed signals with part of frequencies from original. * class weights can also be manually specified.
  • train_w_CM.py: Based on train.py, adding a paralled second branch for taking the static connectivity matrix as one extra input.
  • train_test_PA.py: [Not used in the mentioned paper] Using the point attention (PA) module. A second network will predict a binary mask to select a subset of sampled points from the orginal input.
  • train_test_FS.py: [Not used in the mentioned paper] Using the band_mask module. A second network will predict a binary mask in the frequency domain and reconstructs signal to-be-classified using selected frequencies before feeding it to the downstream classifier.
  • train_testing_resample.py: [Not used in the mentioned paper] Using the resample module. A second network will produce a new grid (not uniformed, not necessarily the same resolution) to resample the input with interpolation. The resampled signal will then flow to the classifier. It has the option to specify whether all channels will share the same grid or using different grids.
  • train_test_window.py: [Not used in the mentioned paper] Using the window_trunc module. A second network will predict a starting point. Input signals will be truncated from this point with fixed length before going to the classifier. (Does not compile with current implementation, still working on it.)
  • train_DCM.py: [Not used in the mentioned paper] Using the dynamic connectivity matrix as feature, which will be mapped to a latent representation with timely distributed dense layer and then a simply RNN will be used for classification. (Huge gap between train and validation.)
  • train_PT_loc_tune.py: [Not used in the mentioned paper] Script for testing the performance when placing the attention module at different locations.
  • train_clf_vae.py: [Not used in the mentioned paper] Try combining the VAE with current classifier. The current version has problems in the quality of predicted reconstruction, need a fix.