The Classification Research Of Medical Image Data Based On EM Algorithm

In recent years, classification methods in the field of data mining have been widely applied to analyze and study the brain fMRI data. This dissertation conducts a discriminate analysis based on resting fMRI data about 69 patients with schizophrenia and 62 normal people.In order to select significant characteristics, this paper used EM Algorithm,which is different from the past many studies of using LDA, ICA, parameter T test and so on. In this paper, we used the SVM classifer that has been widely applied and has better classification effect. The basic idea is as follows:Firstly, in accordance to AAL Templates, the human brain is divided into go partitions. We transfer the correlation coefficient among these areas into z-value through the Fisher’s r-to-z transform, and in front of the twenties features that are the most significant difference are selected by methods of EM Algorithm, thereby finding the abnormal brain regions. And then, the discriminative features are selected by methods of EM Algorithm, two sample t-test. After that, it uses SVM as classifier, to compare the application of two methods of select the discriminative features the effect of the classification. Finally, Constructing the confidence interval by permutation test estimates the reliability of the classification results.The results shows that abnormal brain regions of the schizophrenia are located in default network, attention network, sensory network and auditory network. These brain areas include inferior frontal gyrus, Heschl, Gyrus, Cuneus, Anterior cingulate and paracingulate gyri, Superior frontal gyrus, Thalamus and so on. These brain regions correspond to the significant difference between the brain areas of patients and normal human, which provides a theoretical basis on pathological study of depression patients and also has practical significance to assist doctors of medical clinical diagnosis. The highest accuracy classification with two sample t-test for feature selection is 74.8%. The highest accuracy of classification with EM algorithm for feature selection is 76.3%.