Classification of brain image data using measures of distributional distance

Methods: We developed a method for classifying spatial distributions of regions of interest (ROIs) in brain images. The proposed method is based on computing the Mahalanobis distance between a new sample and datasets related to each considered class (condition). We predict that the new sample belongs to the class corresponding to the dataset that has the smaller Mahalanobis distance from the given subject. We also compared this method to an alternative classification method based on computing the Kullback-Leibler probabilistic distance between distributions estimated through a non-parametric procedure. After the normalization of image data to a common coordinate system, the proposed method can be applied to both structural and functional brain imaging. Here, we applied it to lesion-deficit analysis and MR datasets, performing classification of realistic brain lesion distributions generated using a lesion-deficit simulator [l]. The spatial statistical model of lesion distributions conformed to the Frontal Lobe Injury in Childhood (FLIC) study [2]. The subjects were classified into two classes according to the subsequent development of attention-deficit hyperactivity disorder (ADHD) after closed head injury. Given a new subject with a set of lesioned voxels, the goal was to determine the more plausible class. In experiments, we varied both the size of datasets for the classes and the number of lesioned voxels belonging to a new subject. For each combination of these parameters, we performed the experiments and monitored the classification performance, measuring accuracy rate as the ratio of the number of hits and the total number of rounds.