Recognition Of EEG And Functional Magnetic Resonance Signals Based On Deep Learning

In the field of biomedicine,deep learning has gradually developed as an effective analytical tool.Electroencephalogram(EEG)and Magnetic Resonance Imaging(MRI)are the most commonly used brain imaging techniques in the biomedical field.In this paper,we use the deep learning method to explore the classification of motor imagery(Motor Imagery,MI)of EEG and the functional magnetic resonance imaging(fMRI)classification of Attention Deficit Hyperactivity Disorder(ADHD).We propose the crosssubject and the multi-center classification model respectively,which breaks through some limitations in current researches and provides a new perspective for understanding the cognitive function of brain.The main contents are as follows.In the MI Brain-computer interface(BCI)system,most existing methods are hard to realize the classification among the subjects,which brings a heavy training burden to the subjects,and obstacles the MI BCI system application and promotion extremely.Therefore,transferring the knowledge of existing subjects to new subjects can effectively alleviate the training burden of the subjects.In this paper,we introduce a suitable and effective learning framework to realize the training free MI BCI systems.Considering that the multi-channel CSP(Common Spatial Pattern,CSP)time series can characterize the activity state of different brain regions in the MI task,we propose a separated channel convolutional neural network,which uses different encoder for different channels to highlight the specificity of brain regions in MI tasks.The encoded features are then concatenated and feed into the recognition network to perform the classification.We used traditional machine learning algorithms as baseline models.Moreover,the quantitative analysis was evaluated on our dataset and the BCI competition IV-2bdataset.The results have shown that the model we proposed can improve the accuracy of EEG based MI classification(2–13% improvement for our dataset and 2–15% improvement for BCI competition IV-2bdataset)in comparison with traditional methods under the training free condition.In the application of fMRI data,most of the existing work is based on data from a single center,and the classification problem of multi-center data is rarely considered,which limits the application of fMRI data.In this paper,based on the previous work,we propose an end-to-end deep learning model for fMRI data to realize a multi-center ADHD classification system.We designed a deep feature extraction network for variable time lengths to obtain abstract features for each brain region.At the same time,the brain region fusion and recombination module based on the attention mechanism was developed to learn the interaction of the activity state of the brain region during the task.As a comparison,the effects of different brain region fusion and recombination modules on classification performance were explored.The models was evaluated on the ADHD-200 competition dataset and compared to the single-center based results in the competition.The results show that the brain region fusion and recombination module can achieve improvement effectively,and the optimal results can be obtained based on the attention mechanism.Our multi-center model with high feasibility has achieved better results than the single-center model.