Classification Of Children With Autism Spectrum Disorders Using FMRI Data And Deep Learning

Autism Spectrum Disorders(ASD)is a neurodevelopmental disorder that occurs in early childhood.The main symptoms of ASD are social and language communication disorders,as well as abnormal interests and behaviors.It is a developmental disorder that occurs before the age of three and is characterized by qualitative disorders of social interaction and communication,with symptoms of behavioral and interest narrowing,behavioral repetition,and stereotyping.At present,most neuropsychiatrists mainly use a combination of interview and behavior inventory to diagnose ASD.These methods are more dependent on the professional quality of doctors and the interview mothod,so that the diagnosis results have strong subjectivity and randomness,and the operation process is time-consuming.Although many achievements have been made in the research of new ASD diagnosis methods,these achievements are still far from clinical application.Meanwhile,the number of children with ASD is on the rise around the world.As a result,the small number of neuropsychiatrists will be under enormous pressure to diagnose the growing number of children with ASD.In order to alleviate this contradiction,promote the early recovery of ASD children,and at the same time promote the development of Computer Aided Diagnosis(CAD)technology for ASD children in China,this paper proposes an objective and highly accurate classification Diagnosis technology for ASD children based on resting state functional Magnetic Resonance Imaging(fMRI)data and Deep Learning(DL).This paper took school children aged 5-13 years as the main research object,employed Independent Component Analysis(ICA)to decompose fMRI data and used the modified Multi-Hidden Layers Stack AutoEncoder(MHLSAE)algorithm to distinguish ASD children and the typically developing children.The method finally achieved high classification accuracy,and the results of the study can provide one of CAD method to neuropsychiatrists for diagnosis of ASD children.At the same time,it can improve the diagnostic efficiency of doctors and relieve the social pressure of doctors.In this paper,the results of MHLSAE algorithm were statistically analyzed by cross validation.This paper discussed the universality between the research object and the classification model from the perspective of comparison of the three classification evaluation indexes(accuracy,sensitivity and specificity).After theoretical analysis and empirical research,this paper came to the following conclusions:1.From the perspective of Functional Connection(FC)and Independent Component(IC),this paper analyzed the feasibility of judging ASD by FC and IC.The fMRI data were selected from an international professional ASD brain image database.The paper chose 156 subjects' fMRI data from New York University(NYU)Langone medical center in the database,then these data were broken down into 50 ICs.According to the relationship between FC and IC,10 key IC characteristic time serials were selected to construct a feature vector,then 156 feature vectors were divided into training set and test set.The Stack AutoEncoder(SAE)model with the softmax classifier was builded for training and testing.The experimental results showed that the accuracy was up to 98.98%(sensitivity up to 98.64%,specificity up to 99.32%)with Cross validation(CV).Therefore,this method successfully evaluated the contribution of the synthesis of 10 key ICs to the judgment of ASD,and proved from the perspective of probability and statistics that the classification method combining the key IC and SAE has great potential to be extended to clinical diagnosis.2.Taking the relationship between the frequency characteristics of cerebral oscillation signals and ASD as the research starting point,this paper proposed to take 84 school-aged children' fMRI from NYU as the research object,used ICA to decompose fMRI,and extracted low-frequency brain signals to construct feature vector.Most of ICs from a subject were selected to constitute the feature vector according to the energy proportion,and then MHLSAE was used for training,testing and classification.According to the low-frequency brain signal energy ratio relationship,20 key cerebral oscillation signals were selected and composed into a feature vector,then 84 feature vectors were input into MHLSAE training,testing and classification.Through CV,87.21% average classification accuracy(89.49% average sensitivity and 83.73% average specificity)was obtained.The experimental results showed that the classification based on the main low-frequency signal and MHLSAE reduced the dependence on ASD medical knowledge and verified the feasibility of CAD.3.In order to further validate MHLSAE universality and improve the classification accuracy of the classification model,this paper chose 198 school-aged children from three different institutions as the research object,then directly extracted time signals of all independent compenents to build feature vectors.the input node of MHLSAE was increased,and the algorithm was used to train the new classification model.Finally the performance of the classification model was verified with the CV,and the classification method based on full brain frequency signals and MHLSAE achieved an average accuracy of 96.26%(average sensitivity of 98.03% and average specificity of 93.62%).The experimental results showed that MHLSAE was more suitable for high-dimensional sample sets with a large number of subjects.At the same time,this method completely eliminated the dependence on medical knowledge,and the average accuracy and average sensitivity were greater than 95%,which improved the feasibility of CAD.This paper proposed three different classification methods according to different research objects,and got better and better classification results.