Research On The Classification Method Of FMRI Data Based On Deep Learning

Functional magnetic resonance imaging(fMRI)is one of the hot topics in neuroimaging.Its principle is using magnetic resonance imaging technology to detect changes in the blood dynamics caused by neuronal activity,and then obtaining a large number of three-dimensional brain imaging data.Classifying the fMRI data can effectively decode the cognitive state of the brain,which is very important to understand the working mechanism of the brain for human.However,the high dimensionality of fMRI data is a great challenge to this study.In recent years,a large number of machine learning methods have been used to solve the problem of fMRI data classification,and have achieved good results.Deep learning,which is a new machine learning method,has become an important method to solve this problem because of its excellent modeling ability for high-dimensional data.However,the deep learning models in the existing research are weak for the local feature extraction of fMRI data,and it can not take full advantage of the timing characteristics of fMRI data to improve the classification performance.To solve above problems,we complete the following two tasks in this paper:(1)To solve the problem that existing classification models can not effectively extract the local features of fMRI data,we present a classification model of fMRI data based on Convolutional Neural Network(CNN).Firstly,it designs a CNN structure,and constructs a Restricted Boltzmann Machine(RBM)model by means of the convolution kernel size.Then,the interested region voxels in fMRI data are used to construct and form input data to pre-train RBM,and the weight matrix obtained is relatively transformed to initialize CNN parameters.Finally,the final classification model is obtained by training the whole model initialized.The results on Haxby and LPD datasets show that this model can effectively improve the classification accuracy of fMRI data.(2)To solve the problem that existing classification models do not take full advantage of the timing characteristics in fMRI data,we present a classification model based on Recurrent Neural Network(RNN)for dealing with time-series fMRI data.Firstly,it trains a Convolutional Neural Network with labeled data,and gets the relevant network parameters.Then,it combines label data with unlabeled data by time series,and inputs them into the previous training model,to extract the features of the fully connected layer.Finally,it composes above features into ordered pairs with the pattern of “one label one time-series”,and obtains the final classification model through training.The results on the Haxby dataset show that the model can get a better classification accuracy by using RNN,and the classification performance of the model has been further improved after adding the unlabled data.