Deep Learning based Lightweight Model for Seizure Detection using Spectrogram Images

Epilepsy is a severe neurological disorder, which is onset by the abrupt and erratic electrical gushing in the neurons. Epileptic seizures can be diagnosed by monitoring the brain’s electrical activity using Electroencephalogram (EEG) signals. Conventionally this analysis was done manually by neurologists and had various limitations, but now it is increasingly being automated to save time, minimize human errors and relieve the neurologists from excessive burden. In this study, the EEG signals are first converted into spectrograms. These spectrograms are then fed into the proposed Convolutional Neural Network (CNN) model to automatically learn the robust features and perform binary classification. The proposed CNN model, containing only 3.94 million parameters, obtained an accuracy of 90.9% and achieved precision, recall, and AUC of 91.1%, 93.5% and 97.9% respectively. This work is extended by applying transfer learning on four pre-trained networks VGG16, ResNet, DenseNet, and Inception using the same dataset. Among all these networks, DenseNet achieves the best performance having an accuracy of 92.6% followed by ResNet with an accuracy of 90.3%, Inception with an accuracy of 88.8%, and VGG16 having an accuracy of 88.5%. Although DenseNet achieves slightly better accuracy than the proposed CNN model, it contains almost twice the parameters (8.1 million) in the base model.