Early Screening System For Autism Based On Deep Learning

Autism spectrum disorder is a widespread developmental disorder caused by neurological disorders,and is often found when children show unusual self-consciousness in early childhood.Most sick child show developmental disorders,accompanied by abnormalities in social skills,communication skills,interests and behavioral patternsAutism spectrum disorders cannot be completely cured by existing medical methods,and can only be relieved by acquired intervention.According to the current treatment experience,before the age of six is the golden period for intervention,but when the existing detection methods are diagnosed,it is far later than this period.In clinical applications,the diagnostic criteria for autism are mainly the scale method in the "Diagnostic and Statistical Manual of Mental Disorders(Fifth Edition)" provided by the American Psychiatric Association,which requires doctors to treat patients’ social behavior,language,movement and repetition.There is no effective and objective diagnostic method.In order to enable the subject to be diagnosed with autism in a timely manner,and to solve the problems of low accuracy and long diagnosis cycle of traditional autism diagnosis methods,this thesis uses deep neural network to design a set of autism magnetic resonance imaging diagnosis system.The proposal of this system completely changed the traditional detection method of autism spectrum disorder and greatly improved the objectivity and accuracy of detection.The system uses computer vision technology to analyze subjects’ magnetic resonance imaging and evaluate their performance on early screening for autism.This thesis is devoted to applying deep learning methods to medicine and designing a screening network with short detection cycle and high detection accuracy to detect autism.In this way,doctors can focus more on treatment and reduce the cost of testing.In cooperation with local tertiary hospitals,we successfully collected magnetic resonance imaging data of children with autism and normal children for comparison.A network structure with high accuracy for f MRI data screening is designed after several attempts.In this experiment,1102 cases of structural magnetic resonance imaging data and 2352 cases of functional magnetic resonance imaging data were collected.In the experiment,the performance of the neural network under two kinds of data was compared and analyzed,and finally a network model with high classification accuracy was designed.Experiments show that the accuracy of our network designed by us is 92.1%,and the disease analysis of a single subject can be completed within 40 minutes.Finally,this thesis uses the network interpretation method to explain the classification results of the network and locate the brain regions of interest to the network.We extracted localized brain information that most likely reflected the physical cause of autism.