But not in the presence of metal.

Magnetic susceptibility can be thought of as the “magnetisability” of a substance. Tissues with differing magnetic susceptibility will have a different static magnetic field (B0) strength within them. Thus, adjacent tissues of differing magnetic susceptibility have microscopic magnetic field gradients between them. As you know from the Larmor equation, this means that the precessional frequency of net magnetisation vectors in those tissues will be different. These tiny gradients cause dephasing within a voxel, which leads to reduced signal. However, the magnetic susceptibility of metal is much higher than that of tissue, such that around metal very large variations in Larmor frequency occur. This not only causes signal reduction, but signal loss. (But lost to where? …we’ll get back to that later.) So called signal pile-ups can also occur due to non-linear frequency-position mapping.

As a result, around metal implants, anatomy can be obscured. How can these unwanted effects be mitigated in MRI?

(Note: All acronyms and abbreviations in this post are explained at revisemri.com/abbrev)

Use Turbo Spin Echo

In gradient-echo based sequences, the transverse magnetisation decays according to T2*, which includes effects from magnetic field variations. Image artefacts due to metal are from the inhomogeneity they cause in the static magnetic field, and so gradient echo based sequences are more prone to metal artefacts (especially EPI). However spin echo based methods use RF refocusing pulses to return the T2* decay of transverse magnetisation to (longer) T2 decay, mitigating signal loss.

Many pulse sequences are based on gradient echo because of its speed, including the first sequence in an MR examination: the survey. It is sensible to use a turbo spin echo based sequence for the survey, to allow better planning of subsequent scans. If possible, turbo spin echo should be the sequence of choice for all the subsequent scans too. However, if gradient-echo based sequences must be used, the tips listed below which are not related to TSE still apply.

Set a shorter TSE echo spacing. This will allow collection of more echoes in an echo train before the signal has decayed away (according to T2 in a TSE sequence). On some scanners, the TE and echo spacing parameters may be decoupled (set independently) by selection of an asymmetric k-space profile order.

Use an intermediate/high number of TSE echoes in each echo train (shot duration about 4*TE).

If, in any T2w TSE scan, you usually employ the Fast Recovery method of driven equilibrium (FR aka DRIVE, RESTORE), it would be wise to turn this off, since the assumption that the turbo spin echo refocusing pulses properly refocus transverse magnetisation (so that the flip-back 90° pulse can return it to the longitudinal axis) is violated around the metal.

No Parallel Imaging

Parallel imaging methods necessarily cause a loss of SNR, which is dependent on the g-factor of the coil in use and the acceleration attempted. Since many of the metal artefact reduction techniques listed here will also cause SNR reduction, further SNR loss is to be avoided.

Parallel imaging techniques which rely on a low resolution coil sensitivity information (SENSE, ASSET, mSENSE) will have erroneous or missing sensitivity information around metal implants. Artefacts which arise from these data (or rather lack of data) can be unpredictable. The artefact power can be high in the centre of the field-of-view (FOV). Whilst the artefact power of k-space based methods (GRAPPA, GEM, ARC) is more likely to be smeared across the field-of-view and therefore may not obscure anatomy as much, they have less SNR than SENSE based methods, and SNR loss is to be avoided.

No Sensitivity-Based Homogeneity Correction

As a corollary to “no parallel imaging”, sensitivity based homogeneity correction is also best avoided (CLEAR, PURE, Prescan Normalize). Artefacts due to erroneous or missing coil sensitivity information from the presence of metal is likely to propagate into the “corrected” image; the process will cause further signal loss where tissue is located.

Increase Receiver Bandwidth

Just as chemical shift artefact in the frequency encoding direction occurs in voxels containing water and fat, due to the different resonant frequencies of the two, geometric distortion arises from “incorrect” Larmor frequencies produced around metal implants. The effect of chemical shift artefact is reduced by increasing the receiver bandwith (rBW). On some scanners, this is achieved by reducing the water-fat shift parameter. This causes the range of resonant frequencies over which the distortion is spread to cover a smaller pixel range, and the in-plane geometric distortion is contained within a smaller area within the FOV.

If your scanner has an option to use a higher gradient performance level, select it. This is because higher receiver bandwidth is achieved (all other things being equal) by employing a higher frequency encoding gradient amplitude. Thus, freedom to reduce geometric distortion will be extended.

An increased receiver bandwidth (rBW) will also allow a shorter echo-spacing in the TSE echo train, and a shorter minimum TE (required for T1w and PDw images).

Note that higher receiver bandwidth causes an SNR loss (SNR ∝ 1/√rBW), because the noise power is increased relative to lower rBW. Therefore, don’t necessarily set the a receiver bandwidth to maximum; use a selection equivalent to a water-fat shift of about 0.5 pixels.

Use Higher Resolution

With the receiver bandwidth set—fixing the water-fat shift in pixels to a specified value—the anatomy over which those pixels extend can be reduced by increasing in-plane resolution. There will be a number of consequences of this. SNR will decrease (fewer protons per voxel), scan time will increase (more phase encoding steps), truncation artefacts will decrease. However, note that since MRI magnets are clinical tools, the prescribed resolution parameters will be preserved if the requested receiver bandwidth is not compatible; check that rBW or WFS haven’t changed when you specify your voxel resolution.

Acquire Thinner Slices

Susceptibility effects not only cause signal loss and distortion in-plane, they also cause slice profiles which deviate from the expected planar sheet. As a result, thicker slices can result in partial volume effects through-plane, which can cause SNR loss. So, contrary to expectation, thinner slices might actually increase SNR around a metal implant. However, SNR will decrease in the rest of the image as normal.

Increase Signal Averages

Since most of these measures cost SNR, an increase in signal averages (NSA, NEX, ACQ) is necessary. This will prolong scan time, but a scan time of 5 minutes ought to be achievable, by trading off with in-plane resolution if necessary.

Fat Sat: Use STIR, not a Spectral Method

Spectral selection (CHESS) based fat saturation methods (including SPIR, SPAIR, “Fat Sat” and SPECIAL) are dependent on good main magnetic field (B0) homogeneity. This is why over large FOVs, even in the absence of metal implants, STIR is sometimes preferred. In the presence of metal implants, the B0 homogeneity is significantly compromised, and CHESS-based methods are compromised. STIR is based on the difference in T1 relaxation times between water and lipid, not their chemical shift. Thus, the false apparent “chemical” shift around the metal does not affect STIR.

More about fat saturation methods may be found in a recent fat suppression methods post.

Other Comments

An in-plane resolution increase has the smallest effect in reducing metal artefact. So if, after following the other tips (including signal averaging and allowing sufficient scan time), you need to sacrifice something, start with in-plane resolution.

In addition to these general metal artefact reduction principles, there may be one or two more esoteric tweaks specific to your manufacturer’s pulse sequence implementations. Your magnet manufacturer’s Applications expert will advise.

Even when the metal artefact reduction strategies listed here are applied, SNR of the resultant image will be lower than a conventional protocol. The radiologist assessing the image must be aware of this unavoidable effect.

The Future…?

We’ve talked about signal loss around a metal implant. However, that signal isn’t completely “lost”; it does go somewhere. The signal from tissue around a metal implant has a different resonant frequency than required, which can correspond to a different location in space. A number of techniques have been reported in the research literature to return some of that signal back into the image. For example, View Angle Tilting (VAT) achieves some in-plane correction, but can result in blurring. Slice Encoding for Metal Artifact Correction (SEMAC) extends VAT with z-encoding to resolve distorted excitation profiles that cause through-plane distortions. Some resolution limitation occurs to keep scan times reasonable, but this technique is promising. A longer method is Multi-Acquisition with Variable Resonances Image Combination (MAVRIC), in which multiple 3D acquisitions are acquired with different frequency offsets, and the resultant range of off-resonance images are summed at each slice location. All of the these methods, and others, are still under active research and development in the research community.

Thanks to Marius van Meel for his excellent MARS talk which was basis of this post.

Let’s review them.
It can get confusing.

Phase Encoding

Generically, phase refers to the difference between two points in the time of a cyclical motion or process. Perhaps the most common use of phase in MRI is in phase encoding. Phase encoding is the introduction of a phase variation between the precession of the net magnetisation of spin isochromats across the field-of-view, the degree of which is changed over a series of signal acquisitions, to simulate frequency encoding (rate-of-change-of-phase is frequency). This encoding allows spatial localisation of MR signals in the phase-encoding direction, using the Fourier transform.

A brief (and dense) snippet of explanation such as this does not do justice to phase encoding, which can be difficult to grasp, but it is covered elsewhere on this site and in every MRI textbook.

Phased array

The term “phased array” originates from development of radio wave transmission. In wave theory, a phased array is a group of antennas in which the relative phases of the respective signals feeding the antennas are varied in such a way that the effective radiation pattern of the array is reinforced in a desired direction and suppressed in undesired directions.

In MRI signal reception, we can use a local receiver coil to get higher SNR, because a smaller coil is sensitive to noise signals from a smaller area of the patient. To get the desired anatomical coverage of a larger area, we use an array of the smaller coils (sometimes called coil elements). To get optimal SNR from a phased array coil, it is necessary to make sure that the noise from coil to coil is largely uncorrelated. Part of achieving this involves ensuring minimal electromagnetic interaction between the coils. So in a similar manner to radio wave transmission, adjusting the receive sensitivity of an array of smaller coils is about the shape and arrangement of those coils. There are several competing factors to be considered in the design of a phased array coil, and determining the optimal arrangement of the coil elements is an area of ongoing development.

Phased array coils are sometimes called multiple-element coils. Sometimes the number of independent receive circuits (or channels) is referred to (which also indicates the number of elements), e.g. a 32 channel array. The number of independent RF receiver channels must match (or be greater) than the number of coil elements used in the receiver coil, unless the scanner’s RF system is independent of the number of coil elements in the receiver coil (pdf) in which case any number of receiver coil elements is compatible, with all coil elements used independently.

Heart Phases

A single heart beat can be divided into multiple equal parts in time, called heart phases. “Phases” in this sense is used in the same sense as the phases of the moon as it waxes and wanes: the division of a cyclical process into equal small parts. In cardiac MRI, a functional “cine” imaging scan may be acquired over a few heartbeats, from which we create a single heart-beat movie. The number of frames in that movie of one beat is the number of heart phases; it is the temporal resolution. More heart phases means more data acquisition and a longer breath hold for the patient, but makes for a smoother movie of the beat. Fewer heart phases means a jerkier movie of the heart beat, and possibly blurred myocardial wall boundaries.

Preparation Phases

When you press Start Scan, the scanner does not immediately begin acquiring data. Instead it makes a few clicks, pops, and buzzes, before the scan starts. These are preparation phases, with which the scanner gathers essential data, and optimisation information for the coming scan. It will: check that the right coil is attached (and that all the channels are working); check for signal correction levels; make sure the receiver coil is receiving at the right frequency; determine the right amount of RF power to be used; make sure the various RF frequencies are levelled; check what the optimum resonant frequency (f0) is now that a patient is in the bore; check patient width; check the difference in the delays of the x, y and z gradient channels; perform B0 shimming (and re-check f0); check what signal will cause clipping and adjust the receiver gain; correct for intensity distortions in echo readout; make a noise measurement; gather data for phase correction…

Not all preparation phases are required for every scan. Additionally, some preparation steps can be skipped if a recently acquired scan is sufficiently similar (and the scanner can automatically re-use some of the results from the preparation phases in the previous scan). In one modern scanner, coil sensitivity data for parallel imaging, and B1 calibration scans (to remove so-called dielectric shading) are made part of the preparation phases and are automatically acquired when they are needed.

There are a number of fat suppression methods. Which one you choose depends on the pros and cons of each technique. These change with field strength, field-of-view size, whether regional or global fat suppression is required, whether an increase in scan time is acceptable, etc. Additionally, the absolute quality of fat suppression may not motivate the choice of technique; contrast between tissues of interest may be more important. Overall SNR in an image may also be a deciding factor.

Here is a brief summary of fat suppression techniques.

Short inversion-Time Inversion Recovery

Short inversion-Time Inversion Recovery (STIR) employs a 180° inversion pulse to invert all magnetisation. Then imaging proceeds after a delay, when the longitudinal recovery of fat magnetisation has reached the null point, when there is no fat magnetisation to flip into the x-y plane. Tissues with a T1 relaxation time different to fat have a signal, because they either have not yet reached the null point, or have recovered past it. Most tissues recover more slowly than fat, and so a STIR images have intrinsically lower SNR. Care has to be taken in interpretation of contrast between tissues because of the incomplete relaxation of the water signal of tissues when the image is acquired.

STIR is often preferred when spectrally-selective techniques may not be ideal (large fields-of-view, lower field strengths, areas of high magnetic susceptibility), and the necessary inclusion of the inversion time (TI) to null fat increases scan time. Note that STIR is based on the difference in T1 relaxation times between water and lipid, not their chemical shift. If the inversion pulse is is adiabatic, STIR also becomes insensitive to B1 inhomogeneity.

Spectrally-Selective RF Pulses

RF pulses are tailored to excite protons in a particular resonant frequency range. This range can be narrowed so that the RF pulse affects only water, or only fat (unlike STIR, where all magnetisation gets inverted). This works better at higher field strengths where these resonant frequencies are more separated. Good magnet (B0) homogeneity is required to make this frequency-selective excitation effective, and so techniques based on frequency-selective excitation are more effective over smaller fields-of-view. In general, this technique is called CHESS (CHEmical Shift Selective). If an excitation pulse is water-only, fat may be considered “suppressed” by dint of it being left alone.

A fat-selective CHESS RF pulse can be used as a preparation pulse. After a delay, when the longitudinal recovery of fat magnetisation passes throught the null point, MR image acquisition can occur such that minimal signal from fat contributes to the image. This technique is called SPECIAL (SPECtral Inversion At Lipid). The RF preparation-pulse angle can be reduced to closer to 90° so that the inversion time is as short as possible, which saves imaging time. In this way the preparation pulse is more like a saturation pulse (and “inversion” is a slight misnomer). This is called SPIR (Spectral Presaturation with Inversion Recovery), or on some systems simply “Fat SAT”.

If the RF pulse is adiabatic, making it insensitive to B1 (flip angle) inhomogeneity, a full 180° pulse is used, followed by spoiler gradients which ensure any magnetisation in the transverse plane is dephased. Then MR excitation for data acquisition occurs after a delay (longer than that of SPIR) to allow fat to reach its null point. This is called SPAIR (SPectral Attenuated Inversion Recovery).

Composite RF Pulses

Composite RF pulses can be used to produce a signal from only water protons by making use of the dephasing of fat and water. They are RF pulses made up of a series of shorter RF pulses with small delays between them. They can be quite complicated, but here is a simple example to explain the method.

First, a 45° excitation pulse flips both fat and water. Then after a short time, fat and water are exactly out of phase (both still at 45°, but with opposing transverse components of magnetisation, and thus have a 90° angle between them). Another 45° RF pulse is then applied which flips the fat net magnetisation back to M_z, and puts the water magnetisation in the x-y plane, providing a fat-suppressed signal. This method of fat suppression does not depend on the frequency separation of fat and water (good for low field MR where that separation is small) and is relatively insensitive to B1 non-uniformity (good for high field MR where B1 uniformity is more challenging), but it is sensitive to B0 inhomogeneity. One implementation of this method is called ProSet (PRinciple Of Selective Excitation Technique).

Regional Saturation Bands

Regional saturation employs a 90° RF pulse which, when combined with a gradient orthogonal to the imaging plane, affects only a part of the field-of-view. If imaging follows immediately, no signal will be returned from the suppressed region, since there is no longitudinal magnetisation available to receive the RF excitation pulse. It can be used: to suppress fat within regions of images where the fat signal obscures a tissue of interest; to mitigate aliased signals into a region-of-interest; to reduce the effects of chemical shift displacement of signal in volume selection; or to define the region of interest itself by suppressing surrounding signals (especially in MR spectroscopy). The saturation bands are called REST slabs (REgional Saturation Technique), Presat or SAT bands.

Slice-Selective Gradient Reversal

Slice-selective gradient reversal (SSGR) is possible in spin-echo based sequences, and is appropriate at higher field when chemical shift between fat and water is larger. SSGR relies on through-plane chemical shift being in opposite directions for the 90° and the 180° pulses, so that the shifted fat doesn’t receive both RF pulses and therefore no spin echo is formed from the fat. This is achieved by inverting the polarity of the slice selection gradient associated with the 180° refocusing pulse. SSGR is effective over large fields-of-view and may be combined with other methods of fat suppression above.

DIXON-based

Dixon’s method relies on acquiring images at carefully chosen echo times and using pixel-by-pixel image algebra to calculate a “water only” or “fat only” image. DIXON methods differ from the other methods described in this post in that they postpone the water and fat separation until reconstruction. In this way some of the drawbacks of the other methods are avoided.

Here is the basic idea. Two images are acquired, one at a TE when fat and water are in-phase, and another when fat and water are out-of-phase. Then a water-only image can be calculated using (Image1+Image2)/2. In its most basic form the technique is straightforward, but in practice to make it work a number of non-trivial extensions to the basic technique are required both in data acquisition and in the calculation of water-only or fat-only images. This is because the basic method assumes perfect B0 homogeneity (which is not possible in the presence of a patient), complete absence of eddy currents, negligible susceptibility effects, and it does not account for variation in echo amplitudes. The extensions to the basic method account for these false assumptions.

DIXON-based fat suppression can be very effective in areas of high magnetic susceptibility, where other techniques fail. Note that the TEs are usually fixed in order to make the method work, and so it is not an add-on method for other sequences.

Magnetisation transfer based

A recently reported technique relies on magnetisation transfer (MT). A brief recap of MT follows.
As you know, radiofrequency (RF) excitation pulses have to be at the Larmor frequency of the hydrogen atom (¹H): on resonance. What you may not know is that in MRI we use ¹H in free water molecules; other water molecules are around, such as those attached to macromolecules and membranes (we call these ¹H bound or restricted). These other ¹H have a very large range of Larmor frequencies and have such a short T2 relaxation time (less than 1ms, due to their restricted mobility) that they are not visible in MR images. We can excite or saturate some of the bound water protons by applying an RF pulse off-resonance (i.e. not on the resonant frequency of free water). Then the magnetisation of these bound protons is transferred to the free ¹H protons and the free ¹H behave as if they have received some of the off-resonance RF pulse directly. This magnetisation exchange is called magnetisation transfer (MT). MT is usually used to provide another contrast mechanism because the effect of MT varies between tissues; if we saturate the bound/restricted ¹H , varying amounts of saturation occurs in the free ¹H of tissues.

So back to the fat suppression method. It’s a simple image subtraction of an image with presaturation of both tissue protein and membrane phospholipid protons, from an image without presaturation. In the with-saturation image, efficient MT between water and tissue protein and membrane phospholipid means water gets saturated too, yielding an almost fat-only image. Subtract this from a regular image and you get a water-only image. The nice thing about this method is that fat signals are removed irrespective of their chemical shift (of which there is a range in vivo). It’s also not affected by B0 or B1 inhomogeneity. There is a small reduction of signal from water because not all the water gets saturated via MT, though in most tissues the eventual water signal-loss is small. A downside of the technique is a twofold increase in scan time, and possible misregistration between the two images before subtraction.

Have I missed a method? Comment below!

The value of higher field strength for clinical imaging has been indicated in some clinical applications. Research studies are likely to confirm clinical utility of 3.0T vs. 1.5T—studies showing not just signal-to-noise ratio, contrast-to-noise ratio, or even diagnostic sensitivity and specificity—but effects on patient management, and ultimately, effects on patient outcome.

What are the MR physics issues which are relevant when comparing field strengths?

SNR (signal-to-noise ratio) goes up
… but perhaps not as much as you might think.

Ignoring relaxation effects, the MR signal induced in a receiver coil is proportional to the square of the magnetic field (B0). However, the noise has a linear B0 dependence at field strengths greater than 1.0T, and as a result, SNR is linear with B0 in this range. One might expect, therefore, to double the SNR when the field strength is doubled.

But this theoretical SNR increase is not normally realised in vivo. In practice, SNR gains of this order of magnitude are only realised in certain tissues (e.g. cerebrospinal fluid). At 3.0T, susceptibility effects are more significant than at 1.5T; microscopic susceptibility changes cause larger local magnetic field gradients and greater dephasing of spin isochromats results, decreasing the apparent T2 and causing faster signal decay. For example, the value of the apparent T2 in grey brain matter is likely to be predominantly determined by iron concentration (and thus scales with B0). In grey and white brain matter, due to increases in T1 (see later) and due to the low levels of iron, the SNR gain at 3.0T compared to 1.5T is only 30-60%, not 100%. Furthermore, a higher receiver bandwidth is often used to reduce the larger chemical shift seen at 3.0T. Higher receiver bandwidth reduces SNR.

However, the increase in SNR is still probably the most significant consequence of going to 3.0T affecting clinical utility. SNR increases may be traded for higher spatial and/or temporal resolution, if desired, which makes for demonstrably better image quality. One signal average (1 NSA or on some systems, NEX) at 3.0T should yield a SNR comparable or better than 2 averages at 1.5T (changing from one to two averages produces a √2 increase (41%) in signal).

Other techniques, such as the Blood Oxygen Level Dependent (BOLD) susceptibility effect used in functional MRI (fMRI), also benefit from increased SNR at 3.0T. The BOLD effect is very small, and higher SNR allows higher sensitivity to the BOLD effect.

Parallel imaging techniques also benefit. All parallel imaging methods require a sacrifice of SNR, because they allow image creation from fewer acquired data samples. Increased SNR at 3.0T mitigates this SNR loss from parallel imaging.

T1 relaxation times get longer
…requiring sequence parameter changes.

This may seem counterintuitive—you might at first think that a stronger magnetic field would “pull” the net magnetisation vector of any spin isochromat back to alignment with the external magnetic field more quickly—but this classical picture does not help us here. In fact the T1s usually get longer—slower regrowth of the net magnetisation vector in the z-direction. This has to do with the number of resonant protons which are available to transfer energy to the “lattice”, which depends on field strength. You can read more about this in the main section of ReviseMRI.com.

Longer T1 times of tissues means that pulse sequence parameter settings from lower field strengths may not simply be copied over to a 3.0T magnet. The slower recovery of longitudinal magnetisation usually means that a longer TR is required to maintain expected contrast between tissues. This change in TR has consequences on other parameters and metrics such as scan time and coverage. Similarly, preparation-pulse delay times require modification.

Chemical shift increases
…which is both good and bad.

In the frequency-encode direction, the MRI scanner uses the (precessional) frequency of the MR signal to indicate spatial position in the frequency encoding direction. The different electron (i.e. chemical) environments of molecules in which resonant protons reside can shield (or deshield) the external magnetic field. If protons experience changing magnetic fields, their frequency of precession will change (cf. the Larmor equation). This is chemical shift. Since protons in water in organs and muscle resonate at a slightly different frequency than that of protons in lipids (i.e. fat), the MRI scanner will interpret the frequency difference as a spatial (positional) difference, when fat and water signals are in fact from the same voxel. The frequency shift is approximately 3.5 parts-per-million (ppm) which (according to the Larmor equation) is

• 1.5(T)*42.56(MHz T^-1)*3.5*10^-6 = 223 Hz at 1.5T, or
• 3(T)*42.56(MHz T^-1)*3.5*10^-6 = 445 Hz at 3.0T.

The chemical shifts between water and lipids are actually in a range of 3.3 to 3.5 ppm since chemical shifts can also be affected by temperature and pH. Fat and water are in phase immediately after an excitation pulse, but then 1/223 seconds later (4.5 ms) they’re out of phase (for 1.5 T). For 3.0 T, it’s 1/445 seconds after the excitation pulse (2.2 ms). Thus the in-phase and out-of-phase echo times vary according to field strength.

So, what are the consequences of going to 3.0T?
Chemical shift artefact: Increased chemical shift causes increased chemical shift artefact (which occurs in the frequency encoding direction only, except in EPI-based readouts). An increase in receiver bandwidth (rBW) will reduce the artefact (since chemical shift ∝ 1/rBW ), but with a sacrifice of SNR (because SNR ∝ 1/√rBW). On some scanners rBW can be increased directly, on others in can be increased by decreasing the water-fat shift (WFS) value.

Phase cancellation (black boundary) artefact, which occurs in both the frequency and the phase encoding directions, will occur at different echo times compared to 1.5 T:

• 1.5 T
  • in-phase TEs: 0, 4.5, 9.0, 13.5… (ms)
  • out-of-phase TEs: 0, 2.2, 6.7, 11.2… (ms)
• 3.0 T
  • in-phase TEs: 0, 2.3, 4.5, 6.7… (ms)
  • out-of-phase TEs: 0, 1.1, 3.4, 5.6… (ms)

Spectral fat suppression is more effective at 3.0T because the water resonant peak and the (main) fat resonant peak are more separated. Applying an RF saturation pulse with a transmit bandwidth covering the fat resonance peak only is more easily achieved. A similar argument may be made for water-only excitation. However, note that at large fields-of-view (FOV), conventional (STIR-type) fat suppression is more efficient, because a larger FOV contains a larger B0 inhomogeneity range, and so a spectrally-selective RF pulse doesn’t work so well.
In addition, slice-selective gradient reversal techniques (SSGR) become feasible for fat suppression in spin-echo based pulsed sequences.

Spectroscopy is more effective at 3.0T, due to the greater separation of spectra of different resonant species (choline, creatine, lactate etc), and because of higher SNR. Smaller voxel sizes are achievable, decreasing partial volume effects.

Magnetohydrodynamic effects increase
…but you can forget about them.

When a conductor moves within a magnetic field (B), an electric potential (V) is generated across the conductor. This effect occurs within moving tissue and within flowing blood, most notably in the aorta. The effect of the voltage produced across a vessel containing flowing blood is the magnetohydrodynamic effect.

At 3.0T the consequence of the magnetohydrodynamic effect is similar, but greater in magnitude, to the consequence at 1.5T; the electrocardiogram (ECG) trace becomes non-diagnostic because of an artificially elevated T-wave. However, the vectorcardiogram (VCG) adequately solves the problem of improper triggering from the elevated T-wave instead of the QRS peak, at both 1.5T and at 3.0T.

Even if the magnetic field strength were as high as 4.0T, the voltage generated by the magnetohydrodynamic effect would still be limited to below 40mV (approximate threshold for cardiac depolarisation). (Use, for example, vessel diameter d=1.6cm, average velocity v=42cm/s, in the equation V = dvBsinθ.) Theoretically, the magnetohydrodynamic effect could retard flowing blood, and produce a rise in blood pressure, but the flow reduction would be at most a few percent at field strengths as high as 5.0T.

Dielectric effects (so-called) increase
…causing signal loss, unless you have a RF-transmit system with multiple fully-independent sources

Signal uniformity problems have been observed on conventional 3.0T MR systems, particularly in applications such as breast imaging, imaging of large patients of certain shapes, patients with ascites, and can be observed to a lesser extent in many other imaging applications. The “shading” artefact which is seen comes primarily from a standing wave effect in which travelling waves from multiple coils/elements interfere. These multiple elements are the rungs of the integrated (birdcage) body coil, which is normally used for RF transmission. As a result, a non-uniform B1 field exists in the body. This means the flip angle varies across the anatomy, and signal variations are the result. Dielectric resonance also plays a small part, in which a wave interferes with its reflection from a boundary, but high physiological electrical conductivity levels ensure the role of dielelectric resonance is minor (though it may be observed in phantoms).

A solution to this prominent shading artefact is to independently control the transmit elements; a combination of different B1 fields allows adjustment of the overall B1 field in the patient. This is called RF shimming, and requires multiple independent transmit sources. Four degrees of freedom for each source (waveform, frequency, amplitude, phase) allow a lot of flexibility in obtaining optimum B1 uniformity. RF shimming should be performed on a per-patient basis, and generally reduces SAR, which can be used to enable faster scanning when protocols are SAR-limited.

SAR (specific [energy] absorption rate) goes up
…which can restrict some sequences.

The energy required to tip spin isochromats is negligible compared to the energy that simply dissipates as heat. The International Electrotechnical Commission (IEC) has issued guidelines for safe MRI, to reduce the risk of thermoregulatory distress or local tissue damage. Limits are stated in Watts of RF power per kilogram of tissue. These limits impose a specific absorption rate of RF energy, to limit heating effects. Separate limits are stated for the whole body, and averaged over the head, and in any one gram of tissue. In the clinical range of magnetic field strengths (0.2T to 3.0T), each doubling of B0 produces a four-fold increase in SAR.

SAR limitations can necessitate longer TR times, or poorer coverage, or longer RF pulse durations, or lower flip angles, or some combination of the above. However, a number of SAR management features are applied in recent MRI magnets, in order to maintain pulse sequence parameters and image quality. These include: optimised body coil design; a priori knowledge of energy deposition throughout the body; anatomy-specific dynamic SAR limits; independent RF sensing hardware for feedback control; automatic protocol optimisation for each patient; parallel imaging to reduce the number of RF excitations; modulated refocusing pulses (flip angle sweeps) in turbo spin-echo echo-trains; and most recently, multiple-transmit RF-system achitecture which generally reduces energy deposition hotspots (which are often the limiting factor).

Attraction and torque (and Lenz effect) forces increase
…the usual safety procedures are followed.

Attraction is the pulling force that draws ferromagnetic objects into the bore of the MR magnet. It can make ferromagnetic objects into projectiles, which can produce injury or death to a patient in the scanner bore. Even a paper clip has a terminal velocity within the bore of 40mph at 1.5T, and 60mph at 3.0T.

There are no special consequences for 3.0T compared with 1.5T; continued strict and careful management of the MR unit should minimise associated risk. Positive, documented evidence of safety and/or compatibility of all equipment and devices for the field strength used must be obtained as usual, and the implementation of MR safety should be documented.

Torque is the twisting force which tries to align a ferromagnetic object along magnetic field lines. It is at a maximum at the centre of the imaging volume. It is significant for materials of high magnetic susceptibility, e.g. ferromagnetic materials. Torque is largely shape dependent, and may be more significant than the attractive force. For example, a 1cm needle shaped object will experience a twisting force up to 90 times the attractive force.

Implant contraindications may be more restrictive at 3.0T, and testing is required at that field strength. As usual, positive documented evidence must be obtained that an implant is safe for the field strength used.

The Lenz effect describes a force opposing the motion of an electrical conductor moving in a magnetic field. It may be significant for certain patients with artifical heart valves. You can read more about the Lenz effect in a previous blog-post.
Faster magnetic field changes cause a stronger Lenz effect. Thus, careful observation of at-risk patients whilst moving them into a stronger magnetic field is prudent.

Considering occupational exposure
…an increase in mild, transient sensory effects may occur, no evidence of long term effects is reported.

Long term effects.There is no evidence for cumulative or long-term effects of exposure to magnetic fields up to 4T. Time-averaged static-field exposure limits are not likely to be exceeded; in fact exposure is more like 100 times below recommended exposure limits set in the UK, which are based on ICNIRP guidelines.

Short term effects. Movement within the magnetic field at 3.0T (as opposed to 1.5T) may yield an increase in mild, transient sensory effects such as vertigo, nausea, magnetophosphenes, and taste sensations. Magnetic-field related vertigo results from both magnetic susceptibility differences between vestibular organs and surrounding fluid, and induced currents acting on the vestibular hair cells. Interestingly, it has been shown that the perception of dizziness is not necessarily related to a high value of the rate of change of magnetic field. Magnetophosphenes are not a practical problem for MRI since they are rarely reported for normal MRI exposures even up to 7T. Perception of metallic taste (the electrogustatory effect) depends on direction and rate of head motion, and the threshold for perception of metallic taste varies from one person to the next (and does not depend on the presence of metallic tooth-fillings).

The threshold for minor changes in heart rate, blood pressure changes, and induction of ectopic heart beats, is thought (by the World Health Organisation) to be in excess of 8T. Any effect observed at 3.0T is within the range of normal physiology.

Apologies to the California Milk Processor Board

A one-page cheat sheet is at the end.

What is diffusion weighting?

We use magnetic field gradients to do useful things like encoding. But they also cause dephasing of signal, which we don’t want to happen (this is because when a gradient is on, there is a range of precessional frequencies along the gradient, even within a voxel). Diffusion weighting uses this dephasing effect to our advantage, to show where diffusion occurs.

How does diffusion weighting work?

Simple. A gradient is switched on—a big one to cause lots of dephasing. Then another gradient is used to completely undo all of the dephasing caused by the first gradient. We should end up with no effect, right? Right—but only if tissues and fluids are stationary. If there is motion—including microscopic diffusion of water molecules—then the dephasing caused by the first gradient is not “undone” by the second because the water molecules experienced different a gradient strength from the first gradient to the second, because they moved. As a result, the dephasing stays and we get signal loss on a diffusion weighted image.

So what are b values?

b values are actually a neat way of summing up into one parameter how much dephasing we are going to allow (size of diffusion weighting gradients etc). We need images with different b values to be able to work out diffusion-related parameters such as the ADC (see below). Note that we can’t say what diffusion is from the signal intensity alone, but from the signal intensity loss, which is why we usually acquire a b=0 (baseline) image to compare with.

So what is DTI?

We can only see that good diffusion exists (because of signal loss) along the direction in which a gradient is applied. So if we want to know the diffusion in all directions, we have to get many diffusion weighted images with diffusion weighting gradients in different directions. Ideally “all directions” would mean every possible direction on a sphere, but in practice we do, say, 12, 16, or 32 gradient directions (or more). The actual choice is up to the you. The minimum number of directions we can get away with is six—for example, one diffusion weighting anteriorly, one posteriorly, one superiorly, one inferiorly, one to the right and one to the left. That just about covers 3D space. Of course the diffusion in the brain is not always going to be exactly along one of these directions, so that’s why more directions are often used.

Showing DTI Data

Diffusion Tensor Imaging (DTI) collects information from all the diffusion weighted images (in however many directions was chosen) and tries to sum up all that information about where water can “diffuse to” in each voxel. DTI uses an ellipsoid (a stretched-out sphere, if you like) to represent where water can go. A long thin ellipsoid means very good diffusion for water along the long axis of that ellipsoid. A sphere (not an ellipsoid any more) means the same diffusivity in all directions. The mathematical way of describing the ellipsoid for each voxel is the tensor.

Sorry, what? A tensor?

The tensor is the maths part of DTI. If you like, just think of the diffusion ellipsoid when someone refers to the diffusion tensor. However, it is useful to know that the parametric maps which we produce in DTI (which are images where the pixel values represent some parameter other than signal intensity) are derived from the maths that are used to describe the tensor/ellipsoid at each voxel.

DTI Parametric Maps

It would be nice to draw little 3D ellipsoids at each pixel location, right? Unfortunately that wouldn’t make clinically readable images! So a number of parameters are used which relate to diffusion. There is a choice because the best one for every clinical situation isn’t yet determined. So you can choose. Some examples of parametric maps are now discussed.
Isotropic Image:
If we simply average all the diffusion weighted images which were acquired in all the directions, an image is produced which gives some sense of the total diffusion taking into account all directions. But this Isotropic Image is not generally used clinically, because the ADC is a better map of average diffusion, because the Isotropic Image is affected by T2 shine-through.
ADC: Apparent Diffusion Coefficient (or Constant)
The ADC map shows the average diffusion-freedom water molecules have in each voxel. This parameter can be calculated from all the diffusion weighted images which are acquired in DTI. Note that it is an average of all directions acquired, when performing DTI. The “A” for Apparent is there because the ADC is affected by partial volume averaging, perfusion, and some measurement errors. The average ADC map is sometimes called the trace map, which is to do with the mathematics of how it is calculated. It is not the same as the Isotropic Image; the ADC uses the mathematics of the tensor (the sum of the scalar values of the eigenvalues of the tensor, divided by three (which is the trace/3), sometimes called simply the “trace” image, but let’s not get into the maths now). The ADC is a useful DTI map.
On an ADC map, good diffusion is bright. This is opposite to the diffusion weighted images, where good diffusion is dark. The ADC removes the effect of T2 shine-through.
eADC: enhanced (or exponential) Apparent Diffusion Coefficient (or Constant)
The eADC shows the attenuation of the signal due to diffusion. On an eADC map, good diffusion is dark, just like the diffusion weighted source images. This is opposite to the ADC. Like the ADC, the eADC also removes the effect of T2 shine-through. Clinicians can choose to use ADC or eADC maps depending on whether they want contrast to match (or be opposite to) the diffusion weighted source images.
FA: Fractional Anisotropy
“Anisotropy” refers to how restricted diffusion is. An = not; iso = the same; tropic = direction (from Greek tropos “turn”). So anisotropy means “not the same in all directions”, which is what we are trying to find out about the diffusion of water molecules in each voxel. The ADC and eADC just communicate information about the diffusion in a voxel, whereas anisotropy maps go one step further and communicate information about the orientation of the underlying structure of the fiber tracts in the brain.
There are a number of ways of calculating (and thus, describing) anisotropy. FA is the main one. FA (and RA and VR, below) are rotationally invariant, which is important, because it means that the FA values produced wouldn’t be different if all your diffusion weighting gradients were rotated a bit, or if the patient was in a different position.
RA: Relative Anisotropy
RA is similar to FA, but it is a slightly different calculation (like FA it uses the scalar values from the tensor eigenvectors, but never mind about that now).
FA gives better detail. Use FA.
VR: Volume Ratio
VR is another calculated measure of anisotropy. The SNR and detail of VR is lower than FA and RA. The one thing VR has going for it is that the contrast between regions of low and high anisotropy is stronger than FA or RA.

After the using the Fourier transform to transform our measured k-space data into image space, the image data is of complex type. This image data is then manipulated for different clinical utility. For example, a magnitude image is used to maximise the signal-to-noise ratio (SNR). Phase images are used to measure flow. Let’s look at how our MR signal is recorded and how these image types are calculated.

The changing magnetic field which is the source of the signal measured in MRI is a vector which we represent using complex notation. This is quadrature detection, which refers to the detection of a circularly polarised magnetic field, and results in two data streams with a 90° phase difference. The digitised values from these signals become the real part and the imaginary part of each complex data point in k-space. The ultimate purpose of quadrature detection is to increase SNR by a factor of √2.

Now, the two signals are the real and imaginary channels which are sometimes denoted I (for in phase, the “real” data) and Q (for quadrature phase, the “imaginary” data). The imaginary data is not imaginary in the colloquial sense; it is a measured quantity. These signals are corrupted by “white” noise, which has a Gaussian probability distribution. After the inverse Fourier transform of the complex data, the noise in the complex image data is still white (Gaussian). However, we don’t generally work with the real or imaginary components of the image data (i.e. calculate images using only the real data, or images using only the imaginary data). To use both parts of the complex data values, we calculate magnitude images and phase images, which have physical meaning (proton density and flow, respectively, ignoring contrast weighting and background phase variation for the moment).

Magnitude images are the real and the imaginary parts combined, calculated after the Fourier transform as √(Real²+Imag²), for the complex data point at each image pixel.

Phase images are calculated after the Fourier transform as tan^-1(Imag/Real) for the complex data point at each image pixel. Phase is also known as the complex argument of a complex number.

After making the calculation of a magnitude image, the noise probability distribution is no longer white, and becomes Rician (tending to a Rayleigh distribution as the SNR goes to zero). This sounds concerning, since most MR quality assurance (QA) programs rely on SNR as the primary parameter of choice as a daily-QA metric. However, it seems that if the SNR is above a very low value (<2), the noise probability distribution is approximately Gaussian again anyway. And note that an MR image with an SNR of 2 would not be practically useful.

*Apologies to Philip Mazzei.

Always use these:

• precessional frequencies
• spatial frequencies

Precessional frequencies are the speed with which the magnetic direction of a hydrogen proton (or spin isochromat) rotates around the direction of the magnetic field it experiences. We talk about precessional frequencies, when discussing what’s physically happing to our sample, and when we’re referring to the Larmor equation.

Spatial frequencies are real waves in actual space. Draw a wave with your hand in front of you. Congratulations, a spatial frequency. Now draw a wave with a smaller wavelength. Voilà, a higher spatial frequency. We talk about spatial frequencies when we’re talking about k-space, the signal from the sample, and encoding.

So, precessional frequency relates to position. K-space is a map of spatial frequencies, which are what we record in an MRI experiment. Precessional frequencies refer to individual parts of the sample or patient. Spatial frequencies affect the whole image, and come from the whole sample/patient.

It gets a bit more complicated.

When talking about precessional frequencies, we should state whether we are referring to the magnetic direction of individual hydrogen protons (“spins”), or microscopic groups of spins which experience the same magnetic field strength: spin isochromats. This is because the magnetic direction of a single hydrogen proton wanders around slowly (and precesses quickly) all the time. There is a tendency for the wandering of each spin to be more with the main external magnetic field (B0) than against it, and so overall, a net magnetisation forms. This net magnetisation—of many spins—we call Mz. Because Mz is the sum of many spins, it doesn’t “wander” like the individual spins. It doesn’t precess around the external magnetic field either, unless an RF pulse moves it away from it’s default position aligned with B0; Mz is equal to the equilibrium value M0 when no RF pulse has been applied (thermal equilibrium). We do not talk of “Mz” or “M0″ for a single spin.

So, if you see pictures of magnetisation vectors being manipulated by RF pulses, it should be referring to the net magnetisation of a large number of spins: a spin isochromat.

Read more on the net magnetisation page of the main site.

“…and other off-resonance effects.”

It turns out, off-resonance effects are not a black art or higher plane of MR knowledge after all. Here’s a quick recap. What are off-resonance effects? Off-resonance effects are any signal which has a frequency different from that expected for a nucleus in an idealised system with perfectly uniform static magnetic field (B₀) throughout the sample, and perfect linear gradients. Even if you had a perfect magnet, you’d still get signals which aren’t at the expected (Larmor) precessional frequency.

Since MRI relies on the pulse sequence to relate spatial position in the scanner to precessional frequency, a number of things can occur:

• spatial distortion (images can be distorted globally, or spatial offsets such as chemical shift artefact)
• signal loss (in which net magnetisation vectors within a voxel become out of phase and destructively interfere)
• ghosts
• blurring (in certain k-space trajectories)
• local signal artefacts (dark lines, high signal hot spots).

So you have your perfect magnet, costing you a bazillion groats, or whatever. Proud as Punch, you put a patient in the scanner bore. What happens? The differring magnetic susceptibilities of tissues (think “magnetisability”) causes magnetic field gradients at the interfaces between tissues, and between tissues and free space. These are magnetic field gradients which are not part of our pulse sequence, and cause signal loss, and signals at the “wrong” resonant frequency (off-resonance). These static susceptibility effects are the most common source of off-resonance effects.

Other sources of off-resonance effects are:

• Eddy currents. These arise in conducting structures in the MR scanner when changes in the amplitude of the applied magnetic field gradients occur (i.e. when you switch gradients on and off). Eddy currents decay away, and so produce time-varying off-resonance effects.
• Concomitant gradients (Maxwell terms). Magnetic field changes orthogonal to an applied magnetic field gradient occur due to Maxwell’s equations. So for example, if you apply a z gradient, spatial variations in B_x and B_y also occur. These cause spatial variations in the Larmor frequency and produce off-resonance effects.

Use of spin echoes reduces off-resonant effects. Higher performance hardware can help reduce off-resonance effects, as can the use of parallel imaging in some circumstances. Other strategies include phase correction (using a measured or inferred field map), slice-by-slice shimming, and post-processing corrections such as image registration or image consistency checks.

Thanks to Prof. Jo Hajnal, who gave an excellent introduction to off-resonance effects at ISMRM 2001 (Glasgow).