Classification Of Autism Spectrum Disorder Based On Brain Functional Connectivity And Deep Learning

At the NPC and CPPCC conferences in 2018,National People's Congress representative Wang Xinhui mentioned that children with autism need more care.More and more attention has been paid to the diagnosis and prognosis of autism.In early time,researchers have discovered that there is an obvious abnormality in the brains of autism spectrum disorder(ASD)and typically developing individuals(TD),but now,no specific structural differences have not yet been reached.At present,doctors use DSM to subjectively determine whether individuals have autism tendencies,but this method cannot be used as a diagnostic basis.Therefore,some new methods are needed to help diagnose autism.Along with the rise of functional magnetic resonance imaging(f MRI),it provides a new tool for the study of autism.Resting state functional magnetic resonance imaging(rs-f MRI)is a high-resolution image obtained by measuring blood oxygenation level dependent(BOLD)in the resting state.Therefore,the whole brain correlation analysis is used for the functional connectivity of the brain.The correlation of the time series is obtained by the fluctuation of the BOLD signal in different brain regions in the resting state,and the brain function connection correlation matrix is extracted.Deep learning based stacked autoencoder(SAE)is used to extract features,which are fed to Softmax classifier to distinguish ASD from TD.The main work of this paper is as follows:(1)Analysis of brain function connections from rs-f MRI data.These data are from the ABIDE website and preprocessed for normalization,and then they are further analyzed by the brain function connection to obtain the corresponding correlation matrix,and the relevant elements of the correlation matrix are selected for feature selection.(2)The ASD group and the TD group were categorized into the dataset.The SAE with deep learning was used for feature selection and extraction.These features were finally fed into Softmax classifier,which outputs classification labels.In addition,the cross validation method was used in experiment.For the ASD and TD classifications,it was found that the use of leave-one-out cross validation resulted in an accuracy of up to 95.27%.(3)In order to weak the effect of over-fitting of the ASD classification system,the Dropout algorithm was also added.At the same time,the robustness of the network model in the absence of individual connection information was also enhanced,which further improved the accuracy.This dissertation uses a method based on brain functional connectivity and deep learning to classify autism,uses layer-by-layer analysis of SAE to extract features,and through unsupervised pre-training,supervised fine-tuning,and Dropout algorithms.The network effectively solves local optimal difficulties in training,and makes feature extraction and selection more universal.Through several experiments,it was found that the method not only improved the accuracy of classification,but also provided a reference for the computer-aided diagnosis of ASD,and hoped to provide assistance for clinical medical diagnosis and prognosis.