A New Method For Assisted Diagnosis Of Autism Based On Label Distribution Learning

Autism spectrum disorder(ASD)is a common brain disease.In recent years,with the development of computer science and medical image analysis technology,the use of Resting-state Functional Magnetic Resonance Imaging(rs-fMRI)to diagnose autism has become a hot research topic.The rs-fMRI can effectively reflect the functional changes of the patient’s brain,through the blood oxygen dependent level(BOLD)under non-invasive conditions.Through the medical image analysis of rs-fMRI,we can diagnose autism in a non-invasive and less subjective way.At present,there are three problems with the auxiliary diagnosis of autism: the data set has label noise;the data is high-dimensional and the class is unbalanced;the existing auxiliary diagnosis of autism is mostly a binary classfication,which cannot meet the multi-classification task of ASD in actual tasks.Therefore,we introduce label distribution learning.Through the classifier characteristics of the label distribution learning itself,the problem of label noise is solved and the multi-classification task is realized.At the same time,some useful mechanisms are introduced to deal with the high dimensionality and class imbalance of the data,and finally realize the auxiliary diagnosis of autism.However,using LDL to assist in the diagnosis of ASD still has the following two problems to be solved: one is that the label distribution learning only solves the label noise and multi-classification problems,and other methods need to be introduced to solve the class imbalance;the second is to make some innovations based on the characteristics of the data set based on the label distribution learning to improve the performance of classifier.According to the characteristics of data of ASD,this paper proposes two new methods for assisted diagnosis of ASD based on label distribution learning.Compared with the existing research methods for autism detection,the proposed algorithm is more accurate and robust.The specific contents of the two main tasks proposed in this paper are summarized as follows:1)The first work is to propose a cost-sensitive label distribution support vector regression(CS-LDSVR).The algorithm first uses the DPARSFA tool to process rs-fMRI images and extract the average time series of the brain.After calculating the Pearson correlation coefficient,the brain functional connectivity is obtained;then the classifier of the label distribution support vector regression is constructed,and the cost-sensitive mechanism is added to the classifier.Experiments show that the new method improves the accuracy of the diagnosis of ASD,and compared with the traditional label distribution learning method,the new method overcomes the influence of the majority and minority classes on the results,and can effectively solve the problem of ASD diagnosis.The problem of unbalanced data improves the diagnostic accuracy of minority categories.2)The second work is to propose a label distribution learning with class-shared and class-specific constraints(LDL-CSCS).The algorithm first uses the DPARSFA tool to process rs-fMRI images,extracts the average time series of the brain,and calculates the Pearson correlation coefficient to obtain the brain functional connectivity then the Synthetic Minority Over-sampling Technique(SMOTE)is introduced to generate minority samples for auxiliary training to overcome the problem of class imbalance;finally,the LDL-CSCS classifier is constructed.The classifier combines c class-shared and class-specific features on the basis of label distribution learning,and improves the discrimination ability of the classifier through the common features between different classes and the unique features of each class.Experiments show that the label distribution predicted by the new method is closer to the true label distribution,and the standard deviation is relatively flat,which shows that the model is more robust than the traditional label distribution learning,and at the same time,the diagnostic accuracy rate has also been improved.