A CORRELATION FRAMEWORK FOR FUNCTIONAL MRI DATA ANALYSIS Ola Friman, Magnus Borga, Peter Lundberg † and Hans Knutsson Department of Biomedical Engineering Depts. of Radiation Physics and Diagnostic Radiology † Link¨ oping University Sweden olafr@imt.liu.se ABSTRACT A correlation framework for detecting brain activity in func- tional MRI data is presented. In this framework, a novel method based on canonical correlation analysis follows as a natural extension of established analysis methods. The new method shows very good detection performance. This is demonstrated by localizing brain areas which control fin- ger movements and areas which are involved in numerical mental calculation. 1. INTRODUCTION The functional processes of the human brain are still poorly understood although much effort has been focussed on re- vealing its secrets. A relatively new and promising tool for this purpose is functional magnetic resonance imaging (fMRI). The purpose of fMRI is to map sensor, motor and cognitive functions to specific areas in the brain. For exam- ple, one might be interested in which brain areas that are ac- tivated by a simple motor task such as flexing the fingers, or in higher cognitive functions such as areas for language pro- cessing or mental mathematical calculations. The physical basis of the method is that oxygenated and deoxygenated blood have different magnetic properties, a difference that can be measured in an MR-scanner. More specifically, the signal intensity in a T2 * -weighted MR image of the brain depends slightly on the local oxygenation level of the blood. This is called the blood oxygenation level dependent signal, commonly referred to as the BOLD signal. When neurons in the brain are active they consume oxygen. Blood with a higher level of oxygenation is supplied to the neurons to compensate for the increased oxygen consumption. How- ever, the neurons can not utilize all supplied oxygen which results in an excess of oxygen in the venous vessels. Since T2 * -weighted MR images partially reflect the blood oxy- genation it is possible to analyze such images to detect areas of brain activity indirectly by localizing areas of elevated oxygen levels. To determine where elevations in oxygena- tion level occur during task performance baseline images acquired at a resting state are also required. For this reason activity activity activity time rest rest rest Fig. 1. A typical reference timecourse used in fMRI. a reference timecourse is specified, where rest and task per- formance are alternated, see Fig. 1. A volunteer performs a task, such as flexing a finger, inside the MR-scanner ac- cording to the reference timecourse and brain images are acquired simultaneously. The resulting data is a number of image slices of the brain where a timecourse of intensity values is obtained in each pixel, see Fig. 2. In active brain regions the intensity timecourses have a component that fol- lows the reference timecourse due to the BOLD effect. The problem is to detect such pixels in the MR images. It is es- sential to capture the state of the brain at a certain timepoint, and therefore a very fast imaging sequence called echo pla- nar imaging (EPI) is used. Unfortunately the EPI images suffer from low signal to noise ratio which makes the detec- tion of active brain areas difficult. In order to obtain useful images it is not possible to have a sampling period less than 2 s. With a typical number of acquisitions 200, the effective time for an experiment becomes about 7 minutes. In this paper a correlation analysis framework for this particular detection problem is described. As a starting point we use an ordinary correlation method in order to detect ac- tive pixels. Then a method based on multiple correlation is introduced, and finally a newly developed canonical corre- lation method [4] is presented. 2. THEORY 2.1. Ordinary correlation analysis We begin by describing an ordinary correlation analysis ap- proach. At present time this is the most widely used method to detect active pixels in fMRI images. Assume that N ac- quisitions of each image slice are taken at subsequent time-