Magnetic Resonance Imaging (MRI) has become the most prominent of all brain scanning methods during the last 15 years. According to Hornak (2004) there are approximately 10,000 MRI units worldwide, and approximately 75 million MRI scans were performed in 2003. The field of MRI is growing, and its utilization in cognitive neuroscience has led to a new era of knowledge about the human brain. The structure and vast amount of water in the human body is the key to MR imaging. Hydrogen has the simplest atomic structure compared to all other elements, and about 70% of the human body is made up of water, containing two hydrogen atoms and one oxygen atom. It is the hydrogen atoms that are central for the production of MR images. Water, or more specifically, hydrogen atoms, emit a signal that is both detectable and recordable. The large magnetic moment created by the single proton in the nucleus of the atom, makes hydrogen atoms extremely sensitive to magnetic resonance. Hydrogen protons have a characteristic called spin, which creates a magnetic field, called magnetic moment. In the absence of an external magnetic field, the hydrogen protons’ magnetic moments are randomly oriented, which yields a net magnetization of zero. The introduction of an external magnetic field will align the protons either parallel or anti-parallel to the net magnetic field. The high energy protons are strong enough to align themselves against or anti-parallel to the magnetic field, whereas the lower energy protons will align themselves with or parallel to the magnetic field.
When the magnetic field is increased, fewer protons are strong enough to align anti-parallel, and a larger number of protons are always aligned parallel to the magnetic field than anti-parallel. Thus, once the parallel and anti-parallel protons cancel each other out, only the small number of low energy protons left aligned with the magnetic field creates the overall net magnetization of the brain. Thus, only a very small number of protons are responsible for the MRI signal. If one million protons are anti-parallel, about one million and one proton are parallel (Bjørnerud, 1996). The fact that only a tiny fraction of the protons contribute to the MRI signal is a fundamental constraint of the method, but this number will increase as the strength of the magnetic field increases.
The net magnetic moment is detected by a coil placed on the head of the participant. However, since the net moment is constant over time, it will not generate a signal in the coil, which only “reacts” to changing magnetic fields. By exposing the protons to electromagnetic radiation, their net moment is forced away from the magnetic field. Thus, they are flipped from being parallel with the magnetic field to being anti-parallel. The net moment of the protons will then begin to rotate around the magnetic field simultaneously with the flip away from their starting orientation. This will induce a voltage within the receiver coil. This oscillating signal voltage over time is the MR signal. The electromagnetic radiation is sent in short pulses, often called a radio frequency pulse. The term resonance is used because the radio frequency pulse must be of a certain frequency to be able to affect the protons.
We are using the software package Freesurfer, developed at the A. Martinos Center for Biomedical Imaging at Harvard Medical School, for the analysis of the MR-data. We have applied two consecutive mp-rage sequences, to be able to minimize the noise resulting from participants’ movements in the scanner. Figure 6 shows the original MR volume, while figure 7 shows the Freesurfer-converted and motion corrected average of the two consecutive sequences.
After motion correction, the signal intensities are normalized, and we end up with a volume called T1 within the Freesurfer system (see figure 8). The T1 volume constitutes the basis for the rest of the volumetric analyses.
New techniques for processing of MRI recordings are continuously being developed both for functional and morphometric applications. With regard to the latter, the emerging automated or semi-automated techniques for segmentation and analysis of brain structures seem especially promising. One such technique developed by Fischl and colleagues at the A. Martinos Center and Center for Morphometric Analysis at Harvard (Fischl et al., 2002) is employed in several of our studies. After the T1 volume has been skull-stripped, this procedure automatically assigns a neuroanatomical label to each voxel in an MRI volume based on probabilistic information automatically estimated from a manually labeled training set. The manually labeled training set is a result of the validated techniques of the Center for Morphometric Analysis, and the automated technique extracts the information required for automating the segmentation procedure. Since there is a considerable overlap in intensities between different anatomical structures (even cortical gray matter and white matter overlap by more than 12%, Fischl et al., 2002), spatial information is required to disambiguate the classification problem. The classification technique employs a registration procedure that is robust to anatomical variability, including the ventricular enlargement typically associated with neurological diseases and aging.
In many of our studies, the automated training set as used in Fischl et al. (2002) is employed. Briefly, the segmentation is carried out as follows: First, an optimal linear transform is computed that maximizes the likelihood of the input image, given an atlas constructed from manually labeled images. Next, a nonlinear transform is initialized with the linear one, and the image is allowed to further deform to better match the atlas. Finally, a Bayesian segmentation procedure is carried out, and the maximum a posteriori (MAP) estimate of the labelling is computed. The segmentation uses three pieces of information to disambiguate labels: (1) the prior probability of a given tissue class occurring at a specific atlas location, (2) the likelihood of the image given that tissue class, and (3) the probability of the local spatial configuration of labels given the tissue class. This latter term represents a large number of constraints on the space of allowable segmentations, and prohibits label configurations that never occur in the training set (e.g. hippocampus is never anterior to amygdala). The technique has been previously shown to be comparable in accuracy to manual labeling.
For a number of analyses, the volumetric data are regressed on intracranial volume (ICV). ICV is in our studies usually calculated based on proton density- (PD) weighted low-flip angle FLASH scans obtained during the same scanning session as the scans used for automatic labeling. A deformable template procedure, similar to the “Shrink Wrapping” procedure described by Dale and colleagues (Dale et al., 1993, 1999), was used to obtain an estimate of the smooth surface surrounding the intracranial space (containing cerebrum and cerebellum, CSF, meninges, and brainstem to a level immediately below the pons).
More refined calculations of regional thickness across the cortical mantle are also possible. Even though such analyses are not a part of the present thesis, they constitute the basis for much of our ongoing and future research. Figure 9 gives a rough sketch of the reconstruction and analyses of MR-data, ultimately leading to the cortical thickness measures.
Dale, A. M., Fischl, B., Sereno, M. I. (1999). Cortical surface-based analysis I: Segmentation and surface reconstruction. Neuroimage, 9, 179-194.
Dale, A. M., Sereno, M. I. (1993). Improved localization of cortical activity by combining EEG and MEG with MRI cortical surface reconstruction: a linear approach. Journal of Cognitive Neuroscience, 5, 162-176.
Fischl, Fischl, B., Salat, D. H., Busa, E., Albert, M., Dieterich, M., Haselgrove, C., van der Kouwe, A., Killiany, R., Kennedy, D., Klaveness, S., Montillo, A., Makris, N., Rosen,B., & Dale, A. M. (2002). Whole brain segmentation: Automated labeling of neuroanatomical structures in the human brain. Neuron, 33, 341-55.
Diffusion tensor imaging (DTI)
Diffusion properties of water molecules in brain tissue measured by MRI can give distinct information about fiber integrity and connectivity. Use of diffusion sensitive gradients in many directions makes it possible to quantify degree of diffusion anisotropy in a given tissue, usually described in terms of the fractional anisotropy (FA). FA is assumed to be related to the integrity of fiber myelinization. Multi-directional diffusion analysis is referred to as DTI, reflecting the fact that the tissue diffusion properties can be described both in terms of magnitude and direction in 3D space (as a tensor) (see figure).
DTI yields an in vivo metric related to the microstructure of the brain’s white matter (WM), thus making it possible to quantify characteristics of specific fiber tracts connecting brain structures. This is highly relevant, given that even simple cognitive tasks involve a complex interplay between multiple brain areas, and that the integrity of the connections between the areas involved thus may be related to cognitive function. Other scanning techniques (structural MRI, functional MRI, event-related potentials) yield information about brain areas that correlate with cognitive performance. However, it is reason to believe that the connections between the different areas are as important for cognitive function as the areas themselves (Cardenas et al., 2005). Disconnection of cortical circuits by decreased white matter integrity has been proposed as a general mechanism of age-related decline in cognition (Bartzokis et al., 2004), a hypothesis that can only be tested in vivo with DTI.
The organization of brain functions in complex neural networks warrants a study of the integrity of the connections between brain structures. DTI offers the possibility of looking into WM micro-structure, and mapping neuroanatomical pathways based on diffusion correlations represents a promising, novel tool for cognitive and clinical neuroscience. Studies reporting robust correlations between DTI data and cognitive function are now emerging, but the method has too rarely been applied to study cognition. The paradigm relating diffusion and cognitive function will enable investigators to identify the specific WM pathways which regulate individual differences in performance on specific tasks, also in clinical condition. Salat et al., 2005ab have demonstrated reduction in FA in specific brain areas with age, and Tuch, Salat et al. (2005) have correlations between FA and reaction time. Preliminary data from a development project in our lab in Oslo (n = 25, age 8-13 yrs) indicate that FA in specific regions correlate with neuropsychological measures (see fig below). DTI is also sensitive to clinical conditions, e.g have four reported studies found DTI differences between MCIs and controls (Kantarci et al., 2001; Muller et al, 2005; Fellgiebel et al. 2004,2005).