BSc-Thesis-Project

A Deep Learning-Based Approach for Diagnosis of Schizophrenia using EEG brain recordings

A paper has been published based on the results and approaches used in this work, the pre-print could be found here arXiv pre-print

Datasets

A public dataset of SZ EEG recordings was utilized for this project
Dataset, a 16 channel EEG recording dataset of 84 individuals (45 Schizophernic, 39 Healthy Control)
Dataset was divided into 5 second segments of 224x224 spectrogram images of EEG signal

Methods

Two Neural Network architectures were used

Plain CNN VGG-16

A VGG-16 CNN architecture implemented in keras was utilized which obtained 96.3% accuracy on test data (100% on training data)
each 5 second sepctrogram of size 224x224 was fed to this network as training data. this method is based on the work of Aslan, Akin

Hybrid CNN-LSTM

This architecture consists of CNN followed by LSTM cells. inputs of the CNNs are spectrogram segments for one subject, which are fed to LSTM sequentially
This method underperfomrs compared to the previous plain CNN method due to the lack of sufficient training data.

Generative Data Augmentation

Deep Generative Techniques have been used in order to augment the spectrogram dataset and obtain better results.

DCGAN architecture

Variational Autoencoder architecture

The variational autoencoder achieved 0.5% accuracy improvement and 0.14 loss decrease.