Abstract
The mainstream method for treating drug-resistant epilepsy (DRE) is surgical resection of the epileptogenic zone. Non-invasive automatic localization of epileptogenic zone can be used to guide electrode implantation and improve the effectiveness and safety of neurosurgical treatments. Previous researchers have proposed a range of methods for this purpose, but these suffer from limits such as unclear post-operative outcomes, invasiveness, limited data volume, and single DRE type. This study constructed a non-invasive epilepsy localization method, integrating sLORETA source imaging, time-frequency analysis, and Visual Geometry Group (VGG-16) deep learning. Firstly, 16-channel scalp electroencephalogram (EEG) from 25 successfully operated DRE patients were included. Secondly, time-frequency features by short-time Fourier transform (STFT), continuous wavelet transform (CWT), and superlets algorithm were extracted. Finally, the VGG-16 network was applied to automatically locate the epileptogenic zone. All three feature extraction methods achieved significant accuracy on the dataset. Using STFT for processing and combining it with VGG-16 for image classification achieved an average classification accuracy of 80.2% and a channel identification rate of 80.7% for epileptogenic zones. After processing with CWT, the accuracy increased to 81.7% and the epileptogenic zone channel recognition rate increased to 81.4%. After processing with the superlets method, the classification accuracy was further improved to 83.1%, and the epileptogenic zone channel recognition rate was increased to 83.3%. This marks the pioneering proposal of a systematic framework for non-invasive localization to the epileptogenic zone.