Research On Gender Difference By Analysis Of Brain Image Based On Deep Neural Networks

The rapid development of brain imaging technology provides us with a basis to better explore the mysteries of the brain.Structural magnetic resonance imaging(sMRI)can clearly show the structural information of the brain,which gives us the insights into the structure of the brain and the lesions associated with brain structures.Human brain sMRI is three-dimensional data with more than 2 million voxels,which is too big to deal with using general machine learning algorithms.Recently,deep learning has made great breakthrough in many fields,for the deep of the neural network can effectively handle large data and high dimensional data.So,in this paper we proposed two different deep neural networks to extract the gender differences and at the same time,complete the classification.We built a deep 3D convolutional neural network model to classify male and female brains based on sMRI.The data were scanned from Human Connectome Projects(HCP)including 876 healthy adults(491 females).In this paper,we use the method of down sampling to increase the sample size.Meanwhile,we use the five-fold cross validation method to calculate the model,and use the 3D deep learning model to realize the feature extraction and classification.As a result,the accuracy of classification is 94.21%.The results show that the human brain can be categorized into two distinct classes: male brain and female brain,and suggest that it is better to treat men and women separately in the research or treatment of psychiatric disorder.We proposed a deep 3D convolution extreme learning machine network(MCN-ELM)model to further improve the gender classification accuracy of the brain.Firstly,we extracted multi-scale features by three scale multi-layer 3D convolution neural networks(CNN).Different scale network has different network structure and different scale convolution kernel size,and the parameters of CNN are randomly generated then fixed without tuning.The only parameters need to be calculated are the weights which are between hidden layer and the out layer of the ELM.In this part,we use ten-fold cross validation method to calculate the model.Finally,the accuracy of classification reaches 98.06%.Compared to deep learning method,MCN-ELM network has extracted more complete structural information,saved computing resources,accelerated training speed,improved recognition accuracy,and provided a new idea for feature extraction and pattern classification of small sample data.