Research On Intelligent Algorithms For Auxiliary Diagnosis Of Autism Spectrum Disorder Using Brain Functional Connectivity On Resting-state FMRI Data

At present,the recognition and diagnosis of Autism Spectrum Disorder(ASD)are still dependent on the observation at a level of simple symptom and the experience of clinicians.The number of qualified doctors to diagnose ASD in China is not enough to meet the increasing prevalence of ASD.So,it is urgent to study an automatic diagnosis method to assist the clinical diagnosis of ASD.Machine learning is an important approach to realizing the automatic diagnosis.There are some difficulties in current machine learing based ASD classification and recognition.For axample,the classification accuracy attained in small data samples is satisfied,but decreases when used in large data samples or in data across large institutions,i.e.,classification accuracy drops significantly in the context of larger population samples from different sites.In order to address the issues in ASD classification,in this thesis,we studied the key technology of extracting features on the large data samples from the Autism Brain Imaging Data Exchange(ABIDE)dataset,optimized the model of pattern recognition,improved the classification accuracy and the robustness of model,and realized the stability of model across different sites.The main research contents of this thesis are as follows:(1)All subjects' fMRI data in ABIDE I(Part I of the ABIDE dataset)was well preprocessed in order to improve the noise-signal ration of data.So,the high-quality data was obtained for constructing a functional connectivity(FC)network.(2)An FC based algorithm for classifying autism and control using support vector machine-recursive feature elimination with a stratified-N-Fold cross-validation(SVM-RFECV)was proposed.We chose 35 regions of interest(ROI)based on the social motivation hypothesis to construct the FC matrix and used them as the original input features.Compared to LASSO(least absolute shrinkage and selection operator),Elastic-net and SVM-RFE,a new feature selection algorithm named SVM-RFECV was proposed.A multifold cross-validation was carried out on the new feature subset formed by removing the bottom-ranked feature.The optimal feature subset should be the ones on which the SVM could get the highest accuracy.Multiple inner loops were carried out in order to search the stable features.The experimental results demonstrate that the proposed algorithm can not only achieve high classification accuracy on large sample data sets,but also achieve better classification accuracy than similar researches on across-site data sets.(3)An improved algorithm based on convolution neural network(CNN)was proposed for classifying autism and control.We used the automated anatomical labeling(AAL)atlas with 116 brain regions to construct the FC network and obtained the original FC feature vector with 6670 dimensions.Compared to the classical CNN,the two-dimensional input was changed to a one-dimensional input,and the corresponding convolution kernel and pooling filter were also changed to one-dimensional ones.Then,PReLU was adopted as the activation function instead of sigmoid or tanh or ReLU.This avoided gradient disappearance or neuronal death when the model reversed its derivation.And then,considering that the algorithm was based on the TensorFlow platform,the sofltmax was used as classifier at the time of pre-training and replaced with the parameter-adjustable SVM after pre-training.Furthermore,dropout trick and L2 regularization for cross entropy loss function were used to avoid overfitting,and the optimal convolution kernel size and pooling method were determined after comparing the multi-scale convolution kernels and different pooling methods.The experimental results show that the improved CNN has better feature extraction performance and better classification accuracy than similar studies on the large sample data set.(4)In order to further explore the application of deep learning for classifying autism and control,an algorithm combining SVM-RFE with Sparse Auto-Encoder(SAE)was proposed.At first,the top 1000 features were selected from the FC with 6670 dimensions by the SVM-RFE algorithm and the insignificant features and some noise were removed at same time.Secondly,the method of updated parameters was optimized using the Adam algorithm when SAE was back propagating.The Adam algorithm can adjust the learning rate adaptively,eliminate the direction of the large swing amplitude,correct the swing amplitude,make the swing amplitude of each dimension smaller,and make the network converge faster.Thirdly,sigmoid was replaced with ReLU so as to further improve the convergence speed of the network.Then,a stacked Sparse Auto-Encoder(SSAE)was constructed by the optimized SAE and the high-level and complicated features could be learned by the SSAE.Lastly,these sophisticated features were fed into softmax for classifying autism and control.The experimental results demonstrate that the proposed algorithm can provide a better solution for the application problem where the ASD sample size is not rich and the original feature dimension is relatively high,and obtain a better classification accuracy than similar studies on the all data samples from ABIDE I.