Diffuse outlier time series detection technique for functional magnetic resonance imaging

Abstract

We propose a new support vector machine (SVM) based method that improves the time series classi cation in magnetic resonance imaging (fMRI). We exploit the robust anisotropic di usion (RAD) technique to increase the classi cation performance of the one class support vector machine by taking into account the hypothesis of spatial relationship between active voxels. The proposed method was called Di use One Class Support Vector Machine (DOCSVM). DOCSVM method treats activated voxels as outliers and applies one class support vector machine to generate an activation map and RAD to include the neighborhood hypothesis, improving the classi cation and reducing the iteration steps with respect to RADSPM. We give a brief review of the main methods, present receiver operating characteristic (ROC) results and conclude suggesting further research alternatives.