Contextual Detection of Fmri Activations and Multimodal Aspects of Brain Imaging

Functional magnetic resonance imaging (fMRI) is a non-invasive method which can be used to indirectly localize neuronal activations in the human brain. Functional MRI is based on changes in the blood oxygenation level near the activated tissue. In an fMRI experiment, a stimulus is given to a subject or the subject is asked to conduct a physical or cognitive task. During the experiment, a nuclear magnetic resonance signal is measured outside the head, and time series of three-dimensional image volumes are constructed. The object of this thesis is to study the localization of activation regions from the constructed time series as well as multimodal aspects of brain imaging. The localization of activation regions typically consists of the following phases: preprocessing of the four-dimensional spatiotemporal data, computation of a statistic image, and detection of statistically significantly activated regions from the statistic image. The statistic image is a three-dimensional map, which shows the statistical significance of the measured experimental effect at voxel level. The detection and localization of the activated regions can be carried out by segmenting the statistic image into activated and non-activated regions. The segmentation is difficult because the statistic images are often noisy and high specificity requirements are set for the activation localization. In this thesis, a computationally efficient segmentation method has been developed. The method is based on the utilization of contextual information from the 3-D neighborhood of each voxel by using a Markov random field model. The method does not require assumptions about the intensity distribution of the activated voxels. The method has been tested using both simulated and measured fMRI data. The use of contextual information increased the detection rate of weakly activated regions. In the simulation experiments, spatial autocorrelations in the noise term altered overall false-positive rates only little. It was also demonstrated that the developed method preserved spatial resolution better than the commonly used linear spatial filtering. In repeated fMRI experiments, variation in the activated regions obtained by the developed method was about the same as or less than with other widely used methods. In addition to the activation localization, the use of multimodal data, including the comparison of fMRI and magnetoencephalographic (MEG) data, is discussed in this thesis. This thesis also includes multimodal visualization examples created from MEG, single photon emission computed tomography, fMRI and structural magnetic resonance imaging data.