Multi-voxel pattern analysis (MVPA) has been widely used in the object category classification of functional magnetic resonance imaging (fMRI) data. Feature selection is an essential operation in pattern classification. Searchlight, based on information mapping, is one method of feature selection. In contrast with traditional methods based on activation, searchlight has more sensitivity and then provides higher statistical power. In this study, we applied two
different feature selection methods, searchlight and activation, combined with linear support vector machine (SVM)
classifier, to investigate the classification effect in classifying 4-category objects on fMRI data. We found that the
average classification accuracies of searchlight were 0.8095 (house vs. face), 0.7240 (house vs. car), 0.7247 (house vs.
cat), 0.6980 (face vs. car), 0.5982 (face vs. cat) and 0.6860 (car vs. cat). For house vs. car, the average classification
accuracy based on searchlight was better than that based on activation (0.7240 vs. 0.7143). Specially, searchlight method performed better than activation for some subjects. The results showed that object category classification of fMRI data based on information mapping were significantly reliable. Our findings suggest that information mapping can be applied in pattern classification in future work.
Support Vector Machine (SVM) is an accurate pattern recognition method which has been widely used in functional
MRI (fMRI) data classification. Voxel selection is a very important part in classification. In general, voxel selection is
based on brain regions associated with activation caused by different experiment conditions or stimulations. However,
negative blood oxygenation level-dependent responses (deactivation) which have also been found in humans or animals
contribute to the classification of different cognitive tasks. Different from traditional studies which focused merely on
the activation voxel selection methods, our aim is to investigate the deactivation voxel selection methods in the
classification of fMRI data using SVM. In this study, three different voxel selection methods (deactivation, activation,
the combination of deactivation and activation) are applied to decide which voxel is included in SVM classifier with
linear kernel in classifying 4-category objects on fMRI data. The average accuracies of deactivation classification were
73.36%(house vs. face), 60.34%(house vs. car), 60.94%(house vs. cat), 71.43%(face vs. car), 63.17%(face vs. cat)
and 61.61%(car vs. cat). The classification results of deactivation were significantly above the chance level which
implies the deactivation is informative. The accuracies of combination of activation and deactivation method were close
to that of activation method, and it was even better for some representative subjects. These results suggest deactivation
provides useful information in the object category classification on fMRI data and the method of voxel selection based
on both activation and deactivation will be a significant method in classification in the future.