Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 53904 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Rabiee, Hamid Reza; Manzuri, Mohammad Taghi
- Abstract:
- In recent years, the spread of neurological disorders worldwide has been increasing, especially in developing countries. Due to the unknown function, complexity, and high importance of the brain, such disorders have been pervasive, severe, prolonged, and impose enormous costs on the individual, the family, and the community. Thus, increasing the knowledge about the brain and its areas in various activities is too vital and can facilitate the diagnosis and treatment of many different and unknown neuro- logical disorders. Different kinds of research have been done to automatically process and find the active and vital areas in various states and brain activities. The problem with most of these models is the low accuracy and performance, deletion of some information, and the low speed, which is not appropriate due to the high sensitivity of the nervous system’s issues. In this study, we try to provide a model for accurate classification of M/EEG signals into activity classes, based on dynamic graph convolutional neural networks with the possibility of processing signed graphs. This model’s input is the graphs generated from the brain’s signals by applying the proposed regulatory method to determine the order and processing interval. In the suggested model, after the classification phase, to find essential and differential subgraphs from the input graph per each class, the interpretation of the network’s performance using the Occlusion technique and the dynamic adjacency matrix learned during network’s training is used. These subgraphs represent the significant sources and the critical connections between them in each class. The proposed model’s results are evaluated on the sensory attenuation and the synthetic datasets. These results indicate that the proposed model is more accurate, faster, and more generalizable than other existing models. In addition to confirming the pre- previous paper’s assumptions and the raw data, the interpretation results show that the cerebellum 4-5 left is distinctive in the passive mode. This region cannot be deduced separately by checking the signals of the raw data
- Keywords:
- Disambiguation ; Graph Neural Network ; Electroencephalogram Signals Classification ; Brain Signal ; Magnetoencephalography (MEG) ; Source Selection ; Active Brain Regions ; Graph-Based Learning
Digital Object List
محتواي کتاب
Bookmark