An Efficient CADNet for Classification of High-frequency Oscillations in Magnetoencephalography

Epilepsy is a chronic neurological disease, and locating the lesions precisely is crucial to the success of epilepsy surgery. The high-frequency oscillations (HFOs) in magnetoencephalography (MEG) of epileptic patients can be used to detect seizures. Due to the inefficient and error-prone operation of traditional HFOs detection, it is necessary to develop an approach for the detection of HFOs, which can automatically classify HFOs in MEG. In this paper, We proposed a novel deep learning-based CADNet for the classification of HFOs in MEG. First, we preprocessed acquired MEG data by short-time Fourier transform (STFT), and the extracted time-frequency domain information was applied for model training after pictorialism. Then, we captured the features from these images through convolutional neural network combined with multi-head self-attention, all these features were input into Dendrite Net for classification. We evaluated our model on MEG dataset, and the accuracy, precision, recall, and F1-score of the optimized model reached 0.97, 0.98, 0.97, 0.97. We compared the proposed CADNet with other deep learning models, the result demonstrates that our model outperforms others.