Brain activation detection by modified neighborhood one-class SVM on fMRI data

•A modified neighborhood one-class SVM algorithm is proposed to detect brain functional activation on fMRI data.•The algorithm is based on the modified probability distribution assumption and the modified neighborhood consistency hypothesis.•The idea of ReHo and the multivariate RV measure are combined to select the features of each voxel.•Our method is effective for the brain activation detection in the whole brain.

The one-class support vector machine (OC-SVM) is a data-driven machine learning method that has been applied as a novel technique for brain activation detection. Several researchers have obtained positive preliminary results using OC-SVMs. Nevertheless, existing algorithms are either too complicated or oversimplified and their performance needs to be further improved. In this study, a modified neighborhood one-class support vector machine (MNOC-SVM) algorithm is proposed to detect brain functional activation on functional magnetic resonance imaging (fMRI) data. This method is based on two basic assumptions: (a) For task-related fMRI data, time series of only a few voxels are related to a particular functional activity or functional area, and these voxels should be identified as activated voxels, i.e., the outliers. In contrast, for resting-state fMRI data, only a small number of voxels are unrelated to any resting-state functional networks. These voxels should instead be taken as non-activated voxels, i.e., the outliers. (b) Close voxels have similarly activated or non-activated states. To improve detection accuracy, we apply the following features to each voxel: the RV coefficient between each voxel and its 26 neighborhood voxels (or fewer than 26 for voxels on the edge of the brain), a flag for isolated voxels and a flag for isolated areas. For both task-related and resting-state fMRI data, our MNOC-SVM method effectively detects activated functional areas in the whole brain.