Classification of functional brain images with a spatio-temporal dissimilarity map

Abstract

Classification of subjects into predefined groups, such as patient vs. control, based on their functional MRI data is a potentially useful procedure for clinical diagnostic purposes. This paper presents an automated method for classifying subjects into groups based on their functional MRI data. The proposed methodology provides general framework using preprocessed time series for the whole brain volume. Using a training set of two groups of subjects, the new methodology identifies spatio-temporal features that distinguish the groups and uses these features to categorize new subjects. We demonstrate the method using simulations and a clinical application that classifies individuals into schizotypy and control groups.

Introduction

Functional magnetic resonance imaging (fMRI) is a noninvasive neuroimaging technique used to study functional activity in the living human brain. The task-evoked signal in fMRI is based on the blood oxygen level dependence (BOLD) effect (Ogawa et al., 1990). Neuronal activity is measured indirectly by recording hemodynamic changes (Jueptner and Weiller, 1995). Data collected on each subject are four-dimensional, with three spatial dimensions measured over time.

This paper addresses the problem of classifying subjects into groups using both temporal and spatial information in fMRI data, which has the potential to assist in early disease detection and diagnosis. For example, mild cognitive impairment (MCI) is a condition that predicts the onset of Alzheimer’s disease. It has been shown via FDG-PET that MCI has a distinct functional signature that is a more accurate predictor of the disease than neuropsychological tests (Chetelat et al., 2005). Another application, from cognitive neuroscience research, is confirmation of hypothesized group differences in BOLD response for distinct cognitive states.

Developing methods for classification of subjects based on neuroimaging data poses significant challenges. Large data size and small signal intensity change inhibit accurate classification. Extracting robust features representative of both spatial and temporal aspects of the data plays a crucial role in success of classification.

There have been a number of efforts to classify subjects into groups based on functional data. Some methods proceed from activation maps generated by the general linear model. Kontos et al. (2003) introduced an approach that uses space-filling curves for mapping 3D space into a linear domain. Wang et al. (2004) used dimension reduction techniques on the space-filling curves for discriminative pattern discovery. Liow et al. (2000) applied linear discriminant analysis on the principal components of PET data to classify HIV-1 seropositive patients into AIDS dementia complex (ADC) and non-ADC groups. Ford et al. (2003) used the same approach for discriminating patients and controls based on fMRI data for Alzheimer’s disease, schizophrenia, and mild traumatic brain injury.

Some techniques focus their analysis on regions of interest (ROI). Bogorodzki et al. (2005) developed a method based on differences in regional brain activity. As part of feature selection for preselected ROIs, mean time intensity curves for voxels correlated with the stimulus are computed and modeled using a mixture of time shifted Gaussian functions. New subjects are classified based on derived feature vectors from multiple ROIs. Pokrajac et al. (2005) introduced statistical distance and neural-network-based methods for classification of brain image data contingent on measures of dissimilarity between 3D probability distributions of ROIs.

Mitchell et al. (2004) applied machine learning methods to the classification of cognitive states based on fMRI data, a topic closely related to classification of subjects. They investigated several feature selection and classification approaches. Following the same framework, Zhang et al. (2005) applied machine learning techniques to classification of subjects based on activation maps. In both applications, Gaussian Naive Bayes, Support Vector Machines, and k nearest neighbor classifiers were examined.

While these methods are promising, most rely on contrast maps or predefined regions of interest and are not designed for detecting subtle temporal differences. This paper examines the feasibility of working directly with time series either from whole brain volumes or regions of interest suggested by results from previous studies to select voxels that exhibit highly discriminating features (based on a measure of temporal dissimilarity) and assign group membership to new subjects.