Classification of Multiple Interleaved Human Brain Tasks in Functional Magnetic Resonance Imaging
Manel Martínez-Ramón (Universidad Carlos III de Madrid, Spain), Vladimir Koltchinskii (Georgia Institute of Technology, USA), Gregory L. Heileman (University of New Mexico, USA) and Stefan Posse (University of New Mexico, USA)
Copyright: © 2008
Pattern recognition in functional magnetic resource imaging (fMRI) is a novel technique that may lead to a quantity of discovery tools in neuroscience. It is intended to automatically identify differences in distributed neural substrates resulting from cognitive tasks. Previous works in fMRI classification revealed that information is organized in coarse areas in the neural tissues rather than in small neural microstructures. This fact opens a field of study of the functional areas of the brain from the multivariate analysis of the rather coarse images provided by fMRI. Nevertheless, reliable pattern classification is challenging due to the high dimensionality of fMRI data, the small number of available data sets, interindividual differences, and dependence on the acquisition methodology. The application of kernel methods and, in particular, SVMs, to pattern recognition of fMRI is a reasonable approach to deal with these difficulties and has given reasonable results in accuracy and generalization ability. Some of the most relevant fMRI classification studies using SVMs are analyzed in this chapter. All of them were applied in individual subjects using ad hoc techniques to isolate small brain areas in order to reduce the dimensionality of the problem. Some of them included blind techniques for feature selection; others used the previous knowledge of the human brain to isolate the areas in which the information is presumed to lie. Nevertheless, these methods do not explicitly address the dimensionality, small data sets, or cross-subject classification issues. We present an approach to improve multiclass classification across groups of subjects, field strengths, and fMRI methods. We use an approach based on the segmentation of the brain in functional areas using a neuroanatomical atlas, and each map is classified separately using local classifiers. A single multiclass output is applied using an Adaboost aggregation of the classifier’s outputs. This Adaboost combined the region-specific classifiers to achieve improved classification accuracy with respect to conventional techniques without previous ad hoc area or voxel selection.