New deep learning scheme for classifying ImageNet-EEG signals

Background and Aim: Classification of Electroencephalography (EEG) signals for Brain-Computer Interface (BCI) applications is one of the most critical research areas in the health area. In recent years, reading the mind is one of the exciting areas of researches, and most evidence shows that it is possible to decode brain activity through EEG signal data. Methods: In this method, we use ۱۲۰۰۰ EEG signals that recorded while a subject is looking at images of ۴۰ ImageNet object class. The proposed algorithm is a two-path network. On one side, the region level information of EEG signals extracts and use as the input of bidirectional long short-term memory (LSTM) network. On the other hand, EEG signals down-sampled and convert to images using wavelet transform. These images then use as inputs to a convolutional residual network. Finally, the outputs of these two networks concatenate, and final output creates. In the proposed algorithm, both brain region-level information and temporal-frequency information of the signals used. These features have enough discriminative information to classify EEG signals. Also, unlike most methods that do not take into account the dynamics of EEG signals, the proposed method uses these dynamics using LSTM and Residual networks. Results: The objective of this paper is proposing a deep learning algorithm for classifying EEG signals evoked by visual object stimuli. Conclusion: The average accuracy of the proposed network is ۹۹.۹% for training data, ۹۸.۷% for validation data, and ۹۸.۷% for test data. For further studies in the field of cognitive neuroscience, we investigate different brain areas and frequency bands of EEG signals to understand the process of image cognition in the brain. The results show that the proposed method performs much better than existing methods.

کلیدواژه ها:

EEG ، Deep Learning ، Wavelet ، BCI ، LSTM ، ResNet