Semi-supervised Learning for Epileptic Focus Localization Using Deep Convolutional Autoencoder

Epileptic focus localization is an important process for successful epileptic surgery. The intracranial Electroencephalogram (iEEG) recording helps the neurosurgeons to take the right decision in patients undergoing epilepsy surgical procedure by delineating the epileptogenic area. This requires accurate classification of the EEG signals from different brain areas into focal and non-focal groups. In this paper, we propose a semi-supervised learning based system utilizing a deep neural network to accurately analyze the non-stationary and nonlinear EEG signals. The presented method is based on training a deep convolutional autoencoder neural network in an unsupervised manner and then using the pre-trained encoder along with multi-layer perceptron for classification of the EEG signals into focal and non-focal. EEG feature extraction and classification are performed in a single automated system rather than extracting handcrafted features as in the previous work. The proposed approach leverages the convolutional autoencoder by reducing the features' dimension and extracting the discriminative spatio-temporal features from the EEG signals. Experimental results show the ability of the proposed method to localize the epileptogenic area with an average accuracy of 93.2% using ten-fold cross validation strategy which is the highest among the state of the art.