Classification of whole brain fMRI activation patterns

Abstract

Functional magnetic resonance imaging (fMRI) is an imaging technology which is primarily used to perform brain activation studies by measuring neural activity in the brain. It is an interesting question whether patterns of activity in the brain as measured by fMRI can be used to predict the cognitive state of a subject. Researchers successfully employed a discriminative approach by training classifiers on fMRI data to predict the mental state of a subject from distributed activation patterns in the brain. In this thesis, we investigate the utility of feature selection methods in improving the prediction accuracy of classifiers trained on functional neuroimaging data. We explore the use of classification methods in the context of an event related functional neuroimaging experiment where participants viewed images of scenes and predicted whether they would remember each scene in a post-scan recognition-memory test. We view the application of our tool to this memory encoding task as a step toward the development of tools that will enhance human learning. We train support vector machines on functional data to predict participants' performance in the recognition test and compare the classifier's performance with participants' subjective predictions. We show that the classifier achieves better than random predictions and the average accuracy is close to that of the subject's own prediction. Our classification method consists of feature extraction, feature selection and classification parts.

(cont.) We employ a feature extraction method based on the general linear model. We use the t-test and an SVM-based feature ranking method for feature selection. We train a weighted linear support vector machine, which imposes different penalties for misclassification of samples in different groups. We validate our tool on a simple motor task where we demonstrate an average prediction accuracy of over 90%. We show that feature selection significantly improves the classification accuracy compared to training the classifier on all features. In addition, the comparison of the results between the motor and the memory encoding task indicates that the classifier performance depends significantly on the complexity of the mental process of interest.