Research On Brain Functional Network Classification Based On Deep Neural Network

In recent years,the classification of brain diseases by extracted the features of brain functional networks has become a hot research topic.At present,machine learning methods have been able to automatically classify individuals and locate the focal areas of brain diseases.However,due to the influence of human factors,machine learning methods are difficult to obtain ideal classification accuracy.In this study,a dynamic brain functional network was constructed from resting state functional magnetic resonance imaging(functional Magnetic Resonance Imaging,f MRI)data of normal subjects(Normal Control,NC)and patients with early mild cognitive impairment(early Mild Cognitive Impairment,eMCI),and the feature extraction and classification of dynamic brain functional network were studied by using deep neural network.(1)Research on brain functional network classification based on dilated convolutional neural network.The sliding window method was used to construct the dynamic brain functional networks in NC and eMCI groups over time.Based on the traditional convolutional neural network(Convolutional Neural Networks,CNN)feature extraction method,a novel dilated convolutional neural network(Dilated Convolutional Neural Networks,DCNN)is introduced.The brain network was input into CNN models with different structures for training,and the effects of different CNN structures,different expansion coefficients and different window lengths on the classification results were analyzed.The experimental results show that,compared with the traditional CNN,DCNN can obtain the most accurate classification results and the fastest convergence rate,and the threshold,window length and other objective factors have little impact on the classification effect of DCNN.This indicates that the expression ability of features extracted by DCNN is stronger than that of CNN after pooling layer.While increasing receptive field,more brain functional connection information is retained.(2)A study on the conversion and classification of brain functional networks based on network embedding algorithm.Firstly,f MRI was preprocessed,brain was partitioned with standardized templates and the time series were extracted,the correlation was calculated and the threshold was set to construct the dynamic brain functional network.Secondly,the network structure embedding algorithm(Structure Deep Network Embedding,SDNE)is used to convert the brain functional network into image form.Finally,DCNN is used to extract and integrate image features and output classification results.However,the brain functional network is sparse to a certain extent,and the image obtained by using SDNE directly differs greatly from the original data.Therefore,this paper designs a Sparse Structure Deep Network Embedded Autoencoder(Sparse Structure Deep Network Embedding,SSDNE),that is,to add sparse terms to the loss function of the SDNE.SSDNE can avoid the overfitting problem caused by more neurons in the hidden layer than in the input layer.The differences of brain functional network classification results under different conditions were compared and analyzed.Through the above studies,On the one hand,it is verified that the feature extraction method of the dilated convolutional layer of the dilated convolutional neural network can retain more functional connection information than the traditional pooling layer,extracts more effective features,and obtains more accurate classification results.In addition,some brain regions were labeled as biomarkers associated with eMCI,suggesting that changes in the features of these brain regions play a decisive role in eMCI classification.On the other hand,this paper successfully extended the brain functional network to the CNN.The final experimental results verify the effectiveness of this method.