The Research On Brain Network Construction And Neural Mechanism Of Facial Expression Recognition Based On MEG

Expression recognition using automated computer algorithms has become a hot research topic recently with the rapid development of human-computer interaction technologies.To propose expression recognition algorithms closer to human brain operational mechanisms,we must consider human neuron connections and neural mechanisms when recognizing expressions.The brain recognizes expressions at millisecond timescales and new research tools are required to study neural mechanisms for expression recognition at this time scale and precisely locate brain regions associated with the recognition task.Magnetoencephalography(MEG)combines high temporal and spatial resolution,and hence is an ideal tool to study neural mechanisms for facial expression recognition.This thesis combined a method based on prior knowledge and data-driven(PDM)with MEG technology and a brain network analysis model to explore cooperation mechanisms between different brain regions during facial expression recognition.Consequently,we obtained a brain functional network to effectively predict five facial expressions: sadness,fear,surprise,disgust and anger.We also obtained 5 specific brain networks distinguishing each expression from the others.The specific study includes:constructed whole-brain network was based on a MEG dataset for 19 subjects performing facial expression recognition by finding Pearson correlation coefficients between two signals(0.9-2s)from different channels.We used the rank sum test,random forest,and backward selection to choose 34 channel pairs with the most discriminative and representative ability from the initial 36585 channel pairs.The Desikan-Killiany atlas was used with brainstorm software to identify brain regions corresponding to signals collected by the 34 channel pairs,and finally obtain 26 functional connections.Then,based on the existing 34 channel pairs,all data for an expression were then used as positive samples,with the remaining four expression type data separately divided into four parts and randomly combined into 256 negative samples.We used backward selection and support vector machine with 256positive-negative sample combinations for brain network optimization to obtain the five specific brain function networks(sad vs others,fear vs others,surprise vs others,disgust vs others,anger vs others).Experimental results confirmed the final 34 channel pairs selected were representative and discriminatory in predicting the five expressions,and were most relevant to the expression recognition task.According to the 26 functional connections obtained by tracing signal sources,facial expression recognition included temporal,frontal,occipital,and parietal lobe activations,with temporal lobe activity level being highest(32.35%).The number of functional connections associated with the left hemisphere was greater than those associated with the right hemisphere.The 5 specific brain networks contain the same neural basis,but have unique brain functional connections.This research results help to understand human neural mechanisms involved in facial expression recognition,and provide a theoretical neurological basis to build more robust automatic algorithms for expression recognition.