New technic of transfer learning for detecting epilepsy by EfficientNet and DarkNet models

Abstract

Epileptic seizures are one of the most prevalent brain disorders in the world. Electroencephalography (EEG) signal analysis is used to distinguish between normal and epileptic brain activity. To date, automatic diagnosis remains a highly relevant and significant research topic which can help in this task, especially considering that such diagnosis requires a significant amount of time to be carried out by an expert. As a result, the need for an effective seizure approach capable to classify the normal and epileptic brain signal automatically is crucial. In this perspective, this work proposes a deep neural network approach using transfer learning to classify spectrogram images that have been extracted from EEG signals. Initially, spectrogram images have been extracted and used as input to pre-trained models, and a second refinement is performed on certain feature extraction layers that were previously frozen. The EfficientNet and DarkNet networks are used. To overcome the lack of data, data augmentation was also carried out. The proposed work performed excellently, as assessed by multiple metrics, such as the 0.99 accuracy achieved with EfficientNet combined with a support vector machine (SVM) classifier.