Spatial Resolution of fMRI Images
With a solid understanding of the time-course of BOLD fMRI, we will now turn our attention to the spatial and temporal resolution of the resultant images. The fundamental unit of space in fMRI analysis is the voxel, a 3D rectangular/square prism the dimensions of which are dependent upon 1) slice thickness, 2) field of view, and 3) matrix size. Slice thickness is relatively self-explanatory, while field of view refers to the amount of the imaged volume that is within a particular slice, and matrix size is based upon the number of voxels imaged in each dimension of a slice. If the slice thickness and in-plane resolution are equal, the voxels become cubes and the spatial resolution is said to be isotropic.
The size of voxels in a given study depend upon the purpose of that study; larger voxels of 5mm to a side are appropriate for images of the entire brain, while studies of a particular region require smaller voxels of 1-2mm to a side. Typically, fMRI data utilizes voxels 1mm across, but this smaller voxel size and subsequently higher spatial resolution does come at the cost of a reduced MR signal-to-noise ratio and longer acquisition times. The former of these compromises arises due to the fact the BOLD fMRI signal is proportional to the amount of deoxygenated hemoglobin in a voxel and so, if a voxel is decreased in size, the amount of that hemoglobin will also decrease and the strength of the BOLD signal will decline. The second compromise, of longer acquisition times, is due to hardware limitations--smaller voxels means more voxels, and more voxels require more slices to be obtained, which takes more time. The real problem with this is less a matter of time, however, and more a matter of magnetic field degradation, because if the acquisition time is overly prolonged, then increasing T2* decay will lead to image blurring toward the end of the acquisition time. Despite these issues, smaller voxels are often preferred for fMRI because larger voxels contain too many artifacts other than the desired neurons, and can lead to signal from confounding structures such as glial cells and functionally disparate but anatomically proximate neurons.
Another hindrance to high spatial resolution in fMRI is the presence of extravascular BOLD signalling, which blurs the desired intravascular BOLD signal. This is particularly problematic in large vessels, especially veins, because as blood flow to an active neural area increases, it leads to a flushing effect in which the deoxygenated blood is pushed out by the high flow of excess oxygenated blood, leading to large BOLD signals in veins downstream of the desired imaging area. Fortunately, this issue can be partially resolved by one of three previously discussed techniques: initial dip observation, spin-echo sequences, and diffusion-weighted imaging. The initial dip is seen only in the highly local region of active neurons, and so it will be absent in those larger-scale areas of extravascular BOLD signal. Spin-echo sequencing helps because the magnetic field of larger blood vessels changes very slowly, and thus their diffusing water molecules experience only very small magnetic field changes during movement; as such, spin-echo can reverse their loss of phase coherence and reduce the BOLD signal they give rise to. Smaller vessels, such as capillaries that supply the neurons of interest, have greater magnetic field gradients that are not susceptible to reversal by spin-echo, and thus the BOLD signal of smaller vessels is retained while that of larger vessels is eliminated by spin-echo. Finally, diffusion-weight imaging helps in that the higher mobility of intravascular spins in blood flowing through large vessels makes them more sensitive to diffusion-weighting, allowing their measurement and subsequent suppression in the final image. While each of these techniques allows for greater spatial resolution, they do cause notable reductions in BOLD sensitivity; still, the net benefit is clear, and has led to their widespread use among modern neuroscientists.
These efforts at improving spatial resolution are particularly important in considering the loss of such resolution that is often an unfortunate side effect of many post-processing analyses of fMRI data. Spatial smoothing, for instance, is the method of blurring data across adjacent voxels via Gaussian filter in order to improve statistical validity and inter-subject comparisons; it improves generalizability of data, but can increase average voxel size significantly and thus notably decrease spatial resolution. Another technique, normalization, or the transformation of fMRI image data to a standard stereotactic plot, also distorts voxel resolution. A third method, region-of-interest analysis, combines adjacent individual voxels into a single, larger voxel to allow accurate mapping of functional divisions in the brain across many subjects; this improves visualization of function in a general region at the expense of functional differences in distinct subregions of that region. Generally, then, these three techniques sacrifice spatial resolution for enhanced functional resolution.
