These notes were put together after consulting the Siemens Applications Manual and email discussions with Greg Brown, Senior Radiographer at the Royal Adelaide Hospital. The notes were first compiled sometime in 1998. With any luck, attention to principles may provide a degree of longevity to the meaning of the notes. Many thanks to Siemens and the ever wonderful Greg Brown.
Grey values in MRI reflect two components: (a) signal intensity and (b) unavoidable noise. Noise is, in principle, unavoidable. It is caused by: (a) electromagnetic noise in the body due to movement of charged particles and (b) small anomalies in the measurement electronics, which depends on (a) the size of the RF coil and (b) the bandwidth of the pulse sequence. (Large coils have a large measurement field, but low SNR and vice versa. The closer the coil to the object, the stronger the signal – the smaller the volume, the higher the SNR. Wider bandwidths decrease SNR.)
The signal-to-noise ratio (SNR) is a criterion for image quality. The SNR increases in proportion to voxel volume (1/resolution), the square root of the number of acquisitions (NEX), and the square root of the number of scans (phase encodings). SNR decreases with the field of view squared (FOV2). Measuring SNR: record the mean value of a small region of interest (ROI) placed in the most homogeneous area of tissue with high signal intensity (eg, white matter in thalamus). Calculate the standard deviation for the largest possible ROI placed outside the object in the image background (avoid ghosting/aliasing or eye movement artifact regions). The SNR is then:
For example, suppose mean signal is 720 and st.dev. of background is 20, the SNR is 720/20, which is 36.
Greg: Remember NEX will increase SNR but will not affect contrast unless the tissues are being lost in noise (low CNR). Scan time scales directly with NEX and SNR as the square root of NEX.
Good tissue contrast relies on optimal selection of appropriate pulse sequences (spin echo, inversion recovery, gradient echo, turbo sequences and slice profile). Important pulse parameters are TR (repetition time), TE (time to echo), TI (time for inversion), and flip angle. For T1 weighted images it is important to select a good TR or TI and T2 weighted images depend on a good selection of TE. Tissues vary in their T1 and T2 times, which are manipulated in MRI by selection of TR, TI and TE, respectively. Flip angles mainly affect the strength of the signal measured, but also affect the TR/TI/TE parameters. Also, contrast mediums can be used to improve contrast outside of the blood-brain barrier.
A contrast-to-noise ratio (CNR) is a summary of both SNR and contrast. It is the difference in SNR between two relevant tissue types (A and B): CNR = SNRA – SNRB
Resolution is a function of slice thickness, field of view (FOV), and matrix size. The in-plane resolution is a function of FOV / matrix size.
As slice thickness increases, signal intensity and SNR increase, but the spatial resolution decreases perpendicular to the slice. Thinner slices are less susceptible to partial volume effects. The number of slices for a 2D pulse sequences is a factor of the selected TR / minimum TR. (Slice distance is a factor of slice gap / slice thickness. If slice distances are too small, there is cross talk between slices that can affect T1 contrast. This can be reduced by selection of interleaved slices, but interleaved slice acquisition can produce large mean intensity differences between adjacent slices.)
The FOV is the square image area that contains the object of interest to be measured. The smaller the FOV, the higher the resolution and the smaller the voxel size. SNR decreases with the FOV2. For example, decreasing the FOV from 350 to 240 mm drops the SNR by more than 50%. (Smaller FOV required higher gradient strength.)
A FOV of about 256 mm is common for brain imaging.
The image matrix size is the number of rows multiplied by the number of columns. A 128x256 matrix consists of 128 rows and 256 columns. The rows or lines are each scanned separately (phase encoding), while the columns are scanned simultaneously (frequency encoding). The columns usually determine the matrix size.
Matrix size effects scan time, resolution and SNR. The measurement time is a function of the number of scans (phase encoding steps) x TR x number of excitations (NEX). In-plane resolution is a function of FOV / matrix size. For instance, with a FOV of 256 mm and a matrix of 128x128, resolution is 2mm, but a matrix of 256x256 gives a 1mm resolution. The SNR is proportional to the square root of the number of scans (rows) and the voxel volume (relative SNR = change of voxel size * sqrt[change # scans]). For example, given a FOV of 256 mm, a reduction from 256x256 to 128x128 increases voxel size by 4 times and decreases the phase encoding steps by 2; SNR improves by a factor of 4 * 1/sqrt(2) = 2.82, more than double. (Note, to retain the same SNR for a 512x512 matrix over a 256x256 matrix, 8 NEX are required and measurement time increases by 16 because the number of scans is doubled.)
An image matrix of 256x256 is common for high resolution brain imaging, but 128x128 or 64x64 is preferable for functional brain imaging to decrease measurement time.
Area resolution is a factor of the FOV / matrix size. The voxel size reflects the spatial resolution, which is a factor of area resolution and slice thickness ( [ FOV / matrix ] * thickness ). The smaller the voxel, the higher the spatial resolution, but the lower the SNR. A loss of SNR while attaining higher spatial resolution can be offset by an increase in the number of acquisitions (NEX) and a longer TR (but this will alter contrast).
Artifacts are signal intensities that have no relation to the spatial distribution of the tissues being imaged. There are four types of artifacts (based on appearance): (a) edge artifacts (ghosting, chemical shifts, and ringing), (b) distortions, (c) aliasing (wraparound) artifacts, and (d) flow artifacts.
Motion Artifacts (ghosting and smearing): Artifacts often result from involuntary movements (eg, respiration, cardiac motion and blood flow, eye movements and swallowing) and minor subject movements. Motion artifacts appear only in the phase encoding direction and appear as ghosts or smears. Motion artifacts can be flipped 90 degrees by swapping the phase/frequency encoding directions. Flow effects can be reduced by using Gradient Motion Rephasing (GMR) or synchronisation of acquisition with motion rhythms. Increasing the number of acquisitions (NEX).
Chemical Shift Artifacts: ...
SNR – voxel size, acquisitions, field strength, coil, bandwidth
Resolution – slice thickness, FOV, matrix
Contrast – TR, TE, flip angle, contrast media, sequence
Measurement Time – acquisitions (NEX)
Q. Is it better to acquire an MPRAGE twice with 1 NEX or once with 2 NEX, given the possibility of movement artifacts between scans?
A. I would use 2 NEX the difference between those two approaches is that in 2 nex each line of K-space is collected twice sequentially rather than the whole set of k-space being collected then repeated.
Q. What qualitative improvement in S/N or C/N would be seen when going from 2 to 4 NEX? Is 4 NEX overkill?
A. For MPRAGE 2 nex is overkill. You will only see a CNR change if the contrast is lost in noise. SNR is a funny thing, when it is poor you can notice a 10-20 % change, when it is high you can barely notice a 100% change. Typically we compromise with spatial resolution, SNR and time and try to get sufficient SNR so that if we dropped 20% you wouldn't notice. There is an old MR saying that if the image isn't a bit grainy then you either don't have enough spatial resolution or you took too long.
Q. We got little chance of getting grey mater (GM) and pia mater contrast, so the best we can hope for is GM and meninges/CSF?
A. Just a bit more on that. We are trying to get a brain surface model. The brain has an adherent PIA, the skull has adherent Dura and in between is the arachnoid mater. The middle meningeal layer. That layer is the one that bridges the gyral folds, it does involute to a degree but not fully. In any case the segmentation issues is like trying to feel out the surface of the brain, by touch, with your thumbs, you can’t get into the sulci without making your own tracts. At extremely high resolution (maybe 0.2 mm) with T2 contrast, we might just do it, but that is not practical isotropically (scan time could be days).
Q. Are we using an optimal TR to get the best contrast between GM and meninges/CSF? What are the known T1 curves for these tissues? If we already have a TR somewhere in between their T1 times, perhaps we could move it closer to the GM T1 to further decrease signal from the meninges/CSF on the T1 image? Are we using a flip angle that will optimise contrast or imaging time? Perhaps we can optimise for contrast rather than time? Maybe we should use an inversion recovery sequence (is MPRAGE an IR seq - if so, what is an optimal Gm/CSF TI)? For T2 weighted images, do we have a TE to get the optimal contrast between meninges/CSF and GM? What are the respective T2 curves for these tissues?
A. Good points, but when I experimented with the effective TI of the MP-rage sequence I saw very little change in T1 contrast. That's good because I can bring the time down by 2 minutes, but the arachnoid/GM issue depends on resolution far beyond what we are capable of. Flip angle is only used to sample the evolving behaviour of MMz in an MPRAGE sequence, it has a relation to TR and T1 but only to ensure an appropriate signal level. Without the preinversion pulse the mprage sequence has no contrast.
Q. I've noticed in the applications guide that no slice gap will introduce cross-talk and affect the T1 contrast - maybe we need a slice gap that is some % of slice thickness to improve contrast also.
A. The lack of slice gap may have a minor contrast effect on the EPI but I doubt that it affects the susceptibility contrast. In MPRAGE we use a 3D acquisition so there is effectively one excited slice. In the SE axial (low res anatomical) we need to be using the same geometry as the EPI for overlay but there is some loss of T1 contrast. Overall I don't believe it is significant.
Our EPI is interleaved, as is our low res anatomical (T1 SE). I interleave if I use an inter-slice gap of less than or equal to 50%.
From reading the siemens manual 1 NEX would seem to low for isotropic 1 mm imaging, but the MP-RAGE is a 3D acquisition so the signal that we see is derived from the entire imaging volume (approx. 160mm x 256 x 256) also because we use 2 sets of phase encoding (3D) the effect is the same as adding acquisitions (NEX).
I'll briefly compare the signal and noise source volumes for the MP-RAGE T1 whole head to the 2D SE T1 situation. (It will be a crude discussion so physicists please forgive me.)
In the 2D SE situation signal is generated from the slice (256 x 256 x 5 mm = 327680 mm^3 = 0.33 litres) Noise comes from the tissue within the sensitive volume of the coil (about 17 litres). The sequence takes 1024 "views" of that signal (256 phase steps, 4 Acquisitions) to generate an image of good SNR in about 3 minutes. We end up with resolution of 5mm X .9 mm x .9 mm
In the 3D situation we use a non-selective excitation of a volume that includes the entire head. The signal is generated from the tissue within the coil (approx 17 litres), the noise from the same volume. The sequence takes 4096 views of the tissue (phase steps L-R and Phase steps A-P x I acquisition) - resolution 1 x 1 x 1 mm
See the differences? Now GRE section of MPRAGE that generates signal is sampling less of the available signal than the SE routine, but the overwhelming advantage of the 3D acquisition is that it can yield acceptable SNR levels (relatively as strong as the 5 mm SE sequence) in 7 minutes, where the 2D SE approach would require at least 75 minutes (3 x 5 x 5). In reality to get the appropriate coverage we would have to change from axial to sagittal and do 9 interleaved sequences so the figure would be 675 minutes (11.25 hours).
MRI is an interesting juggle, and perhaps this illustrates the advantage of the MP-RAGE sequence quite graphically.
Clinically the MP-RAGE we use has adequate SNR and CNR for grey white discrimination and the CNR is even higher for Brain/CSF. I don't believe the segmentation problem is an issue of CSF/Brain CNR or SNR, it is one of resolution and contrast between the brain surface and the arachnoid meninges.
Extreme T2 would give stronger brain to CSF contrast. I don't have a T2 weight MPRAGE, but I will experiment with a sequence called CISS and let you know. CISS may also allow 1 mm resolution or higher.
When determining adequate resolution for the segmentation, I think you have to address the spatial dimensions of the smallest elements you are trying to resolve. In this case that seems to be the inner folds of adjacent grey matter at the bottom of a sulcus. The situation will be improved by cortical atrophy, CSF separating adjacent GM, and resolution.
The V shape of the CSF/brain border of a standard gyrus presents us with a resolution gauging wedge. The finer we can discriminate, the deeper into the groove the model will display. On top of that we have an overlying tissue of arachnoid mater that bridges the top of the sulcal groove and can be included in the model if not resolved or manually excluded from the segmentation process.
1mm res is a practical limit, but it isn't really going to get us far into a tight sulcus. 2 mm would be worse I guess.
I agree with Darren's thoughts about the resolution for the task of looking at the brain as a gross object, but I was under the impression that you would need to resolve finer details for the segmentation issue. If A. Dales group do it with 2 mm resolution, how have they dealt with the meninges and do we have an example of their surface rendered brain?
DW. We can segment a low-res anatomical (T1 weighting) for this purpose, but an EPI will be largely brain and very little other head tissues. A recent paper reports acquisition of 1.5x1.5 mm in-plane resolution, TE=56ms, TI=1200ms, TR=22sec, 4 NEX. They don't report matrix size or FOV.
GB. So long as it has the same geometry as the EPI fMRI images it would also have similar distortions and that may be a very good thing. Their sequence looks a bit like an inversion prepared EPI for FLAIR which will look a bit like T1. I don't understand 4NEX but I can work on that. It would be similar to using your base fMRI T2 weight images and reversing the grey scale before overlaying the activation maps. Resolution is the issue then. Their 1.5 mm in-plane is achievable, I could dial in a range of slice thicknesses 4-10 mm.
DW. It sounds very useful. The higher in-plane resolution and T1 like weighting would provide a volume that should be easily aligned with an MPRAGE volume. Given the EPI like distortions, we could consider a warp alignment, but this makes combining alignments more difficult, so we will most likely use a linear alignment. (BTW, it seems from the 3x3x3 sample volume you sent a while ago that the distortions have been somewhat reduced now, although we didn't see any slices around the sinuses in that sample.) Andrew and I believe we can then apply the alignment parameters to stats maps and effectively reslice them into the MPRAGE space (or Talaraich space).
I imagine an isotropic volume would be best because it tends to reslice well, but failing that, the best in-plane and smallest slice thickness would be preferable. (Does smaller slice thickness reduce susceptibility artifacts?) Puce (1995) has reported use of a 3.2x3.2mm in-plane and 7mm thick volume (128x64 matrix, FOV 40x20). It seems like 1.5 or 2mm isotropic is not an option, but we can get 3mm isotropic and that would be good. BTW, 2 NEX should be O.K. at that resolution, huh?