EEG Source Estimation Using Deep Prior Without a Subject’s Individual Lead Field

Abstract

Estimating current sources in the brain using an electroencephalogram (EEG) is affected by the accuracy of the lead field, which represents signal propagation from the cortical sources to the scalp. To accurately compute the lead field, one must know the subject’s head structure. However, imaging methods for brain structure require large-scale equipment. In this paper, we propose a novel method of EEG source estimation that does not require the lead field of each subject obtained in advance. The current sources in the brain and the lead field are simultaneously estimated using implicit prior distributions expressed by an untrained convolutional neural network (CNN), namely Deep Prior, and a pre-trained CNN using the lead field of an average subject, respectively. The proposed method requires only a noisy EEG observation and the lead field of the average subject. We showed that the proposed method was more accurate than the conventional methods, and was also as accurate as the Deep Prior-based method with the lead field of each subject.