Loading...

Enhanced Neural Source Localization by Exploiting the Relation between EEG and fMRI

Ataie, Ali | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57524 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Amini, Arash; Ghazizadeh Ehsaei, Ali
  7. Abstract:
  8. Electroencephalography (EEG) and the functional magnetic resonance imaging (fMRI) are the two important non-invasive tools in probing the neural activity of human brain. EEG has a high temporal resolution while suffering a low spatial one. On the other hand, fMRI has a superior spatial resolution while offering a low temporal one. Enhancement of the spatial resolution of EEG source localization methods is of high importance due to multiple reasons such as its lower capital and operational costs compared with fMRI, its mobility and ease of use and extra. But all the EEG source localization methods have been the subject to a major problem, and it is absence of a ground-truth about the real neural sources. The main idea in this thesis is to use the simultaneous fMRI signals as the best estimate of neural sources in order to design and train new, precise and reliable EEG source localization models. To do so, we had to build the best modeling on relation of the signals of the two modalities. Based on the fact that the fMRI BOLD measures some quantity proportional to the energy consumed in each voxel, we hypothesized that the power 2 function must be one the most important nonlinear blocks in this modeling. Based on this idea, we designed a setup for a realistic simulation of the simultaneous EEG-fMRI signals. Using this structure, we managed to produce a big amount of simultaneous EEG-fMRI data needed for training a deep neural network model. We designed a fully-convolutional network to predict the simulated BOLD signals based on the EEG measurements. This resulted to about 70% guaranteed localization precision on the simulated data. In order to offer a model to work with the real data, we designed a set of simultaneous EEG-fMRI experiments consisting of four localizer tasks and an event-related one. By cleaning up these data, we fed it to our deep model in order to fine-tune it to achieve a similar precision on the real data from a novel subject (absent in training data). Moreover, we analyzed the performance of our proposed deep model in predicting the brain activities to various stimuli and related stimuli contrasts, as their overlap with the ground-truth fMRI regions. We used the ratio of the intersection of two regions to their union (the Jaccard ratio) to measure the performance of our model on different novel subjects. While our deep model managed to significantly (p<0.05) make overlaps (range from 6 to 30%) with fMRI regions for the auditorial stimulus and all other contrasts, custom EEG localization methods did not so. The custom localization models could only provide significant overlaps about 14% with fMRI for the simple visual task, while the proposed deep model reached a 32% overlap with BOLD activation. Moreover, another important aspect of the simultaneous EEG-fMRI data acquisition is to achieve a simultaneously high resolution both in temporal and spatial dimensions. One major paradigm in this category is the EEG-informed fMRI analysis. In this paradigm, a preferred EEG signal is selected to regress the BOLD response. By implementing the proposed “power 2” idea on these methods we showed that it strongly influences the previous results. Specifically, re-analyzing a previous simultaneous study, using our EEG-power-informed fMRI analysis, we found the role of the gustatory cortex in the process of making a decision between two food items, which was missed by the conventional method. We further analyzed the fifth task of our simultaneous experiment using this new method. Doing so, revealed the substantial and unique role of the “Caudate tail” in encoding the face value memories in human brains. Previous invasive studies on monkeys had proved the role of caudate tail in encoding the value memories of non-natural fractal objects. So, our experiment and novel analysis extended this result to humans and categories beyond non-natural objects, i.e. the faces
  9. Keywords:
  10. Deep Neural Networks ; Functional Magnetic Resonance Imaging (FMRI) ; Electroencephalography ; Caudate Region ; Brain Source Localization ; Deep Artificial Neural Network ; Value Memory

 Digital Object List

 Bookmark

No TOC