I realized I might have rushed a little with the last entries and some MRI basics remain perhaps unknown to you. So in this post I will address (1) how the machine works? and (2) why on earth do I have to go through the pain of preprocessing my images?
- How the machine works? (check this video out)
A large percentage of our body is made up of water. Our hydrogen protons spin aligned to the earth’s magnetic field (which is more or less 25 gausses). When placed inside an MRI machine, the hydrogen protons align to the axis prompted by the MRI’s magnetic field (300,000 gausses, or 120,000 times the earth’s magnetic field). Then, radio pulses force the proton into aligning to a 90ª or 180ª degrees angle.
When this radio pulse is switched off, our protons go back to is original position while emitting a signal registered by receiver coils. Based on the content of water, and depending on the type of sequence, the different tissues forming our brain will need different times to get to their original position. For instance, in T1-weighted sequences, the first and last tissues to recover from this forced alignment are the cerebrospinal fluid (CSF) and the white matter.
- Preprocessing. Why?
Images come out from the scanner in the shape of 3D-ish JPEG files. In this original form, they are practically incomparable and need a couple of arrangements before conduct any analysis.
First and foremost, it is mandatory to get rid of the skull and other soft-tissues not typically considered as part of the brain (i.e., dura, meninges). This step is commonly referred to as skull-stripping, and it is so obvious that it doesn’t worth discuss it any further.
Second, some brains are bigger than others. This is, you won’t find two identical brains (not even in twins). To reduce inter-subject differences emerging from different head shapes and sizes, one would like to first re-orient the images to make sure they all are in the same coordinates.
Right after this, one has to spatially normalize each brain to a goal-standard image, or a very good looking brain. This is generally referred to as a template. In most cases, this average-like brain is the MNI152 template, which is the mean of 152 brains (well, it is not exactly that… but sort of).
These two steps are commonly known as registration. It involves a series of deformations on the image that can be linear (translating or rotating) and non-linear (zoom and shear) to make the brain fit the template’s coordinates and space.
Third. Inside every pixel, or voxel because of the extra dimension, there’s information about intensity. Again, in T1-weighted images, voxels with low intensity will match those with a greater content of water (e.g., CSF).
However, magnetic fields can induce some artefacts that can lead to what is called field inhomogeneities. Basically, you will spot both brighter and darker areas on the brain which are not related to any medical condition but problems with the scanner. This, naturally, may drive to erroneously estimate the intensity of a voxel and mess the next step.
Bias-field correction computes the global intensity mean of each tissue-compartments to re-scale them and remove extreme values making the difference in intensity between tissues more homogenous.
Fourth. Segmentation. Using the intensity values, each one of the tissue classes is classified. As an example, intensity values below 40 would qualify as CSF, values between 50-90 would be considered as grey matter, and intensity values above 90 would be sorted out as white matter.
That said, voxels falling inside a given compartment will be isolated from the rest of classes to form specific maps. In each map, bright voxels would have positive values above 0, whereas black voxels will present values of 0 (i.e., no intensity, no tissue). In the image below, and from left to right, you will see the grey matter, white matter anc CSF maps. Note how voxels with intensity (i.e., bright ones) overlap the tissue that they’re supposedly representing.
Fifth. Smoothing, in a nutshell, averages voxels with their closest neighbours with a Gaussian kernel, which determines to which extension voxels are considered as neighbours (i.e., 4, 6 or 8 mm radius). This prevents extreme intensity values along the tissue maps.
The direct cost of this is, obviously, a loss in spatial resolution (you’ll see the image blurred). However, in doing so you’re improving the signal-to-noise ratio (SNR) and the spatial overlapping within subjects (overcome different gyrification patterns). Smoothing can be done in both structural and functional images. You can find an example of this below.
That’s all folks. Altogether, MRI preprocessing steps are required because images in its original form are not directly comparable.