Master of Science
Electrophysiological source imaging (ESI) is a widespread and no-invasive technique in neuroscientific research and clinical diagnostics. It provides a well-established and high temporal resolution of source activity and gives the brain signal by analyzing the corresponding EEG signal.
However, it is still a major challenge to deal with the domain shift problem between the datasets of different subjects or sessions in ESI problem. Furthermore, the variable noise included in the EEG signals inevitably influence the accuracy of localization of source activity.
In this paper, we propose a novel denoising autoencoder-based unsupervised domain adaptation (DAE-UDA) algorithm to tackle these problems. To the best of our knowledge, it is the first to solve the domain shift in the ESI problem by using UDA to narrow the discrepancy between different domains. Moreover, we innovatively combine our model with denoising autoencoder (DAE) to remove noise and learn a robust mapping from the noisy EEG signal to the brain activity.
Extensive numerical experiments and the analysis of real EEG datasets demonstrate that DAE-UDA can effectively remove noise and mitigates the domain shift of low-SNR EEG signal. Our model outperforms other classical ESI methods in robustly imaging source activities under a variety of source settings.
Summary for Lay Audience
In our daily lives, some people suffer from several neurological diseases such as epilepsy, ADHD, Alzheimer's, etc. These neurological diseases may be caused by some lesion areas in patients' brain. Thus, if we could find an effective method to locate the exact localization of the areas of the lesions, many mental illnesses could be alleviated or even cured. Electroencephalogram (EEG) is a recording of brain activity that measures electrical activity in our brain using small metal electrodes attached to the surface of the top of the patient's head. Due to the non-invasiveness and convenience of EEG, some researchers want to measure and analyze patients' EEG signals to find the source in the brain that generates the EEG signal, which may be the location of the brain lesion area. However, there are two main challenges that bother doctors and scientists: the noise in EEG signals impacts the analysis, and the probability distributions of EEG signals and the source of the signal in the brain often change.
In this paper, we propose a new deep learning method call UDA-DAE to solve these problems. Our model mainly consists of two parts: unsupervised domain adaptation and denoising autoencoder. The first component unsupervised domain adaptation technique (UDA) can overcome the change in the distribution of EEG and source activity in the brain and make our model more applicable, and the second component denoising autoencoder (DAE) can remove the noise contained in EEG signals. Extensive numerical experiments and the analysis of real EEG datasets demonstrate that UDA-DAE can effectively remove noise and mitigates the change of distribution of noisy EEG signal.
Li, Runze, "Denoising-Based Domain Adaptation Network for EEG Source Imaging" (2023). Electronic Thesis and Dissertation Repository. 9246.
Available for download on Sunday, December 31, 2023