Research On Classification Of ADHD Based On The Feature Extraction In Resting State FMRI

Attention Deficit Hyperactivity Disorder (ADHD) is a kind of common childhood dis-eases caused by mental disorder, the main symptoms are hyperactivity and inattention. In recent years, studies found that in some adults also appeared the same symptoms. Our country and even the world have paid high attention to the diagnosis and treatment of ADHD, because the etiology and pathogenesis of ADHD still remain unclear, which is difficult for clinical diagnosis and treatment. Therefore, looking for the objective ba-sis of diagnosis has been a hotspot but difficulty in the field of brain and cognitive neu-roscience discipline.In the study of brain and cognitive neuroscience, pattern recognition technology plays an important role. It combines with support vector machine (SVM) and artificial intel-ligence technology to do the image processing, classification and time signal of func-tional magnetic resonance imaging. In this thesis the experiment is based on the resting state functional magnetic resonance imaging technology to do the research of ADHD’s classification in the light of feature extraction. The whole brain average time series data are supported by ADHD-200competition. After preprocessing, firstly we divided the whole brain voxels into some small modules in certain size, and then use the data of mask to do mask processing on the brain data to remove some background information which are meaningless. Calculates the average of each small module that are remained, and then do the correlation calculation, the correlation coefficients obtained will be the value of function connection characteristics using for the classification. Before the fea-ture classification, we need do the feature ranking by the method of between class and within class distance as preliminary screening. In order to improve the computational efficiency, this experiment adopted the method of locally linear embedding to do the dimensionality reduction, and obtained a mapping matrix in the low dimensional space. Finally doing the classifying by support vector machine and get the model of classifica-tion, and then using5fold cross validation to test the correct rate of classification which can be used to determine the effect of classification.The results show that the classification effect is good, the highest classification accuracy is85.71%, which means our method of brain partition is better, but we still need to im-proved classification algorithm to raise the accuracy.