Diagnosis Of Autism Based On Community Structure Of Brain Network And Deep Learning On Resting-state Functional MRI

It is widely accepted that the normal brain and autism brain are completely different.However,the detailed differences between them are controversial.Therefore,doctors hardly diagnose autism spectrum disorders in an objective way.In fact,the diagnosis of autism is relatively subjective based on observing behaviors and referring to diagnostic tables instead of objective medical evidence.In this sense,scientists really hope to find an efficient and effective method to identify autism spectrum disorders.Fortunately,with the development of science and technology,numerous novel methods have been applied on autistic diagnosis.These methods are helpful to find a new way to diagnose autism.Scientists these years have used functional magnetic resonance imaging(f MRI)to explore human brain since it allows scientists to observe brain activity by checking high-resolution images noninvasively.Among these methods relative to f MRI,constructing brain functional network is very popular.Brain network is a complex network,which means it possess the properties of complex network.One of the most important properties is community structure.That of course exists in brain network.Community structure can be viewed as a way to find out similar structures in a network since the ultimate objective of detecting community structure is gathering similar nodes into the same community and scattering different nodes into different communities.However,community detection is a NP-hard problem,which means that it is impossible to develop a method that is suitable for all backgrounds.So,it is also difficult to develop a perfect algorithm for detecting community structure in brain network.Deep learning,as one of the efficient methods in machine learning,can extract higher dimensional and more abstract features by imitating human activities.It has been successful in voice recognition and image recognition.In term of classification problems,deep learning method can reach much higher accurate rate.So,it has been used on disease diagnosis.Nevertheless,which kinds of data should be inputted into deep learning classifiers is an important problem.In addition,how to make the best classifier for disease diagnosis is another problem because not all deep learning classifiers are suitable.Focusing on these issues,a new algorithm GAcut(Genetic Algorithm Cut)is proposed for detecting community structure in brain network in order to extract different features between autism spectrum disorder and typical control.Based on that,deep learning classifier is used to classify different subjects.To some extent,an objective method has been developed to autism diagnosis.The main content of this paper is:(1)Use popular methods to preprocess rs-f MRI data and then construct brain network for single subject and a group of subjects.Analyze the influences of each preprocessing steps.(2)Describe new algorithm GAcut and make experiments on many real datasets.The results show that GAcut is efficient.(3)Apply GAcut to detect community structure in autism spectrum disorders and typical controls and prove that brain network possesses community structure.Carefully describe the differences of community structures between autism spectrum disorders and typical controls and analyze the pathogeny that cause those differences.(4)Build NMI statistical matrix to concentrate all information into a low dimensional matrix.Use deep learning classifier to classify disorders and normal people.A large number of experiments show that NMI statistical matrix is very useful to classify subjects accurately and fast.