Design And Implementation Of FMRI Classification System Based On Deep Learning

Functional magnetic resonance imaging(fMRI)is one of the main technologies to study the function of human brain.The brain functional connection(FC)obtained from fMRI is widely used in neuroimaging to study mental diseases,which is expected to provide potential biomarkers for the classification or prediction of brain diseases.Due to the large dimension and small amount of data in brain function connection,common classification algorithms are easy to over fit and have poor recognition effect.In recent years,the rapid development of deep learning has brought a lot of new inspiration to the field of medical image.Generation adversarial network(GAN)has achieved good performance in many fields of classification tasks,but its application in fMRI is relatively small.In this study,an improved generation adversarial network is proposed to generate brain function connection samples.In addition,convolutional neural network(CNN)has also achieved excellent performance in the field of nonimage in recent years.In this paper,a one-dimensional convolutional network is proposed to replace the current deep neural network to classify functional connectivity.The main research contents are as follows:(1)In view of the fact that there are few samples in the fMRI open data set of mental disorders,and the traditional machine learning algorithm is easy to over fit,this paper proposes an improved generation adversarial network(FC-GAN)for brain function connectivity data generation.The model is composed of generating network,discriminating network and constraint network.Each part is composed of multi-layer fully connected neural network,which cooperates with each other to improve the effect of data generation.In this work,the joint training method of variational auto-encoder and generative adversarial network is used.At the same time,a constraint network is proposed to restrict the generation direction of generative network,and the purpose of improving the quality of data generation is realized.(2)Aiming at the problem of poor classification effect and over fitting phenomenon of the current all connected deep neural network in brain function connection classification task,this paper proposes a brain function connection classification algorithm based on one-dimensional convolution network according to the characteristics that brain function connectivity is a kind of vector data.One dimensional convolution stack is used to extract the features of the input data,and attention module is added to the network to calibrate the importance of the feature map between the convolution layers.Compared with the current deep neural network,this method has a significant improvement in classification accuracy,and this method is the first attempt of convolution network in brain functional connection classification.(3)The fMRI image classification system is designed and developed by MATLAB.The system mainly includes two modules: data expansion and classification recognition.The functions of data acquisition,data preprocessing,model training,data expansion,classification recognition and so on are realized.The purpose of classification of the input fMRI image is realized,and the diagnostic accuracy of the disease and the work efficiency of doctors are improved.