Disease classification with support vector machine

Abstract

In the field of diagnostic imaging, discriminations between patients and normal subjects have been widely-studied using some statistical approaches for the purpose of clinical diagnostic support. However, there are not enough studies which proposed to discriminate a disease from others among patients. In this study, we tried to discriminate Alzheimer's disease from Parkinson's one by applying support vector machine (SVM) to the single photon emission computed tomography (SPECT) brain images. First, we compared blood flows between Alzheimer's and Parkinson's diseases by the one-sided t-test at each voxel and extracted the voxels whose p-values were <0.01, which constructed the dataset for this study. Then these voxels were divided into the subsets based on brain functions. Secondly, we applied SVM to the dataset with selecting the subsets by means of forward selection. The accuracy rates were 98% and 86% calculated by leave-one out cross validation and external validation methods, respectively. In addition, the brain regions where remained as significant using forward selection were consistent with the SPECT findings obtained previously.