Classification of functional magnetic resonance imaging data using informative pattern features

The canonical technique for analyzing functional magnetic resonance imaging (fMRI) data, statistical parametric mapping, produces maps of brain locations that are more active during performance of a task than during a control condition. In recent years, there has been increasing awareness of the fact that there is information in the entire pattern of brain activation and not just in saliently active locations. Classifiers have been the tool of choice for capturing this information and used to make predictions ranging from what kind of object a subject is thinking about to what decision they will make. Such classifiers are usually trained on a selection of voxels from the 3D grid that makes up the activation pattern; often this means the best accuracy is obtained using few voxels, from all across the brain, and that different voxels will be chosen in different cross-validation folds, making the classifiers hard to interpret. The increasing commonality of datasets with tens to hundreds of classes makes this problem even more acute. In this paper we introduce a method for identifying informative subsets of adjacent voxels, corresponding to brain patches that distinguish subsets of classes. These patches can then be used to train classifiers for the distinctions they support and used as "pattern features" for a meta-classifier. We show that this method permits classification at a higher accuracy than that obtained with traditional voxel selection, and that the sets of voxels used are more reproducible across cross-validation folds than those identified with voxel selection, and lie in plausible brain locations.