The content of this syllabus was adopted from “Ultra-Low-Field MRI and its combination with MEG“ by Lauri Parkkonen, Risto J. Ilmoniemi, Fa-Hsuan Lin, and Michelle Espy from “Magnetoencephalography: from signals to dynamic cortical networks“, Ed: Selma Supek, Cheryl J. Aine, 2014, Springer press

The availability of large arrays of highly sensitive SQUID magnetometers in modern MEG devices enables one to measure magnetic fields other than those produced by neuronal electrical activity. Perhaps the most promising such possibility is to measure magnetic resonance imaging (MRI) signals. This would bring several benefits such as improved registration of MEG and MRI, improved work flow, structural images with less distortion, and information about the conductivity of the brain.

In principle, combining MEG and MRI is straightforward: simply build a magnet as well as gradient and radio-frequency (RF) coils around an MEG sensor array. The problem is that MEG devices are generally designed to measure femtotesla-level fields and have a dynamic range only up to some tens or hundreds of nanoteslas while in MRI the fields go up to several tesla, i.e., 15 orders of magnitude above the weakest fields measured by SQUIDs. The solution is ultra-low-field MRI: the recordings are performed in a field of about 100 microteslas. If conventional MRI approaches, including tuned inductive receiver coils, would be used at these low fields, the resulting signal-to-noise ratio (SNR) would be very low and the system practically unusable, because the SNR is proportional to the square of the field; the signal at 100 microteslas would be 9 orders of magnitude weaker than at 3 tesla. Fortunately, we can use three methodologies to counteract this problem. First, unlike tuned receivers, the sensitivity of SQUID sensors is independent of frequency; second, we can use pre-polarization techniques to magnetize the sample before the MRI data acquisition; third, we can gain from parallel data acquisition made available by the large number (up to 306 currently) of SQUID sensors in a typical MEG array.

NMR

There is a fundamental difference between ULF and high-field (HF: B > 1 T for this discussion) NMR and MRI. In HF, the process of polarization, magnetization reorientation, spin evolution, and measurement all occur within a single magnetic field provided (typically) by a large superconducting magnet. At ULF, however, these fields may all be different and produced by different magnets. Because the magnetization and thus signal is proportional to the field, one strategy is to use pre-polarization in a higher field (B_p ~ 1 - 200 mT) followed by readout in a much lower measurement field (B_m ~ 1 - 200 μT).

There are also differences in approaches to reorient spin between conventional HF MRI and ULF MRI. In HF MRI, spin reorientation is typically accomplished by a time-varying magnetic field applied at the Larmor frequency, in a direction orthogonal to the direction of main magnetic field. In many ULF-MRI configurations, B_p and B_m are orthogonal. Thus, one can simply begin precession by a rapid (non-adiabatic, dB_p/dt >> γB_m²) shut-off of B_p. In a non-adiabatic process, the magnetization cannot follow the field change, and is left aligned orthogonal to B_m. Precession will then begin automatically, without a B₁ pulse.

Once precession has begun, detection of the magnetization can begin. In HF MRI, the typical scanner strengths are 1.5 or 3 T. This translates to a proton Larmor frequency of 64 or 128 MHz. A tuned induction coil is highly sensitive in this range. However, in ULF MRI precession occurs in the B_m field, typically on the order of 100 μT, corresponding to a Larmor frequency of 4.26 kHz. In HF MRI with a Faraday coil, the signal scales as B². One order arising from the magnetization being proportional to B and the other one for the induced signal being proportional to the Larmor frequency. However, this relationship no longer holds in ULF MRI. While pre-polarization is an approach to improve the former factor (Macovski and Conolly 1993), using an ultra-sensitive detector such as the SQUID is a way to mitigate the latter. SQUIDs, broadband detectors with unsurpassed sensitivity of about 1 fT/ Hz in the frequency range of ULF MRI, are almost two orders more sensitive than a Faraday coil in this regime (Myers et al. 2006; Matlashov 2011). Because the SQUID is also the detector of choice for MEG, the combination of MEG and ULF MRI in a single device becomes obvious. We should, however, mention that getting a SQUID to work in the dynamic environment of MRI, even at ultra-low fields, is quite challenging.

Basic principles of image acquisition

Here we will discuss that in addition to lower polarization leading to lower signal, the lower strength of imaging gradients also poses a unique new challenge for ULF MRI in terms of image acquisition time. To improve the spatial resolution, one either has to increase the gradient or the acquisition (or encoding) time. The gradients in ULF MRI are typically on the order of 10⁴ T/m. This is about 1/100 of those used in HF MRI. Thus, if we want to keep the same resolution as in HF MRI, with these weaker gradients we must increase the acquisition time by a factor of 100.

Why don't we just increase the gradient strength? There are two reasons: the first being related to concomitant fields. These are the unwanted magnetic fields that inevitably arise in directions orthogonal to the measurement field B_m. In HF MRI, these concomitant fiels are neglected, because the main magnetic field (B₀ > 1 T) is typically much higher than the gradients (10^-2 T/m) and thus frequency variations produced by the stray fields are small. However, in ULF MRI the main magnetic field B_m ~ 10^-4 T and the gradients G ~ 10^-4 T/m, which generates a magnetic field variation of 0.2x10^-4 T in a 20-cm FOV. The concomitant fields, not in parallel with B_m, are of similar order of magnitude as G. Thus, the frequency variations are non-negligible and the total magnetic field experienced by the spin system is no longer in a plane orthogonal to B_m. In general, concomitant gradients can be accounted for with some effort (e.g., Volegov et al. 2005; Nieminen et al. 2010; Hsu et al. 2014) but they do pose a constraint.

The other reason that gradients cannot be arbitrarily strong is related to bandwidth. Consider a simple case: a HF MRI system with B₀ = 1 T and 10^-2 T/m gradients. A 20-cm object would have a central frequency of 42.6 MHz, and a frequency spread across the object of Δω ~ 85.2 kHz. However, if B_m = 10^-4 T and G = 10^-4 T/m, the central frequency is 4.26 kHz and the frequency spread within a FOV of 20 cm is 852 Hz. If we turn up the gradients, we will further widen the frequency spread across the object and consequently need to measure part of the MRI signal in the challenging low-frequency regime.

Thus, our only choice to maintain spatial resolution at ULF appears to be longer acquisition times. But at ULF, the T₁ and T₂ times for many interesting tissues are approximately of the length required for ta (Zotev et al. 2009), so we are also running out of signal at the same time.

A simple way to speed up imaging is provided by parallel imaging methods (Pruessmann et al. 1999). In this approach, the spatial sensitivity of an array of SQUIDs is used to replace spatial encoding steps. Using less encoding steps speeds up the acquisition but usually results in an aliased image. By using, for example, the sensitivity maps from an array of coils, a full image may still be reconstructed. In applications like combined MEG and MRI, where a SQUID array is available, this has been demonstrated as a viable approach (Zotev et al. 2008a). Some of the complications due to inductive coupling between tuned receiver coils in HF MRI parallel imaging are greatly reduced with untuned SQUIDs. However, it should be emphasized that parallel imaging accelerates the MRI acquisition at the cost of SNR, which is already scarce in ULF MRI.

ULF-MRI Instrumentation

In the discussion below we will generally assume that the application is for the combination of ULF MRI and MEG. Thus, we are assuming that ULF MRI is being done in the presence of a magnetically shielded room (MSR). We will also assume that we have an array of SQUIDs. We note that for applications of ULF MRI that do not include MEG, an MSR may not be required. For example, the Clarke group at UC Berkeley operates in an aluminum eddy current shield only. We also note that there has been progress using sensors other than SQUIDs (namely the atomic magnetometer) both for MEG (Xia et al. 2006) and ULF MRI (Savukov et al. 2009), which we will not discuss here. In ULF MRI, one aims at the highest pre-polarization that can be tolerated (while maintaining the benefits of the ULF regime) and at the most sensitive sensor. The heart of any ULF MRI instrument is the SQUID sensor array. Large (hundreds of sensors) SQUID arrays have been used for decades for MEG (Hämäläinen et al. 1993). The noise level of a SQUID can be as low as 10^-15 T, enabling it to detect the very weak (10^-12 to 10^-15 T) magnetic fields from brain activity from outside the head. A challenge for ULF MRI is that the changing magnetic fields are many orders of magnitude larger than the dynamic range of SQUIDs. Several strategies have been implemented to deal with this. One approach is to encapsulate the SQUID chips in sealed Pb boxes that have a critical field of about 80 mT. The SQUID can also be locally shielded by Nb plates on the chip (Luomahaara et al. 2011). In either case, the pick-up coil extends outside the shield to detect fields of interest. Thus, the input coil circuit also needs current limiters. One can use externally controlled superconducting cryo-switches (Zotev et al. 2007), which become resistive when heated above the critical temperature and thereby make the pick-up coil much less sensitive to magnetic fields. Other groups have used arrays of Josephson junctions (Hilbert et al. 1985) that become resistive above the Josephson critical current and thereby limit the current.

To achieve a high pre-polarization field, a variety of magnet designs have been proposed. In some sense, generating Bp is much simpler than at HF because the fields are so much lower and the field homogeneity requirement is less stringent (McDermott et al. 2002; Burghoff et al. 2005). B_p can be relatively (a few percent) inhomogeneous compared to the parts-per-million or better requirement for B₀ in HF MRI.

However, there are considerations to producing (and removing) B_p. In fact, some of the advantages (reduced homogeneity requirement, shorter T₁ times) can be disadvantages if not accounted for. The first consideration is that it is not trivial to make a pulsed field at > 50 mT. The coil will heat up, dissipated energy must be removed, and the proximity of a large amount of conductor near the SQUIDs can introduce Johnson noise. The B_p coil should be physically disconnected during the measurement (via a relay) to reduce the antenna effect.

There have been several approaches to producing a pulsed B_p, including water-cooled (Myers 2006), coolant (Fluorinert) and liquid nitrogen (LN) cooled coils (Sims et al. 2010). The LN coil has the benefit of 7× lower resistance, but requires an additional cryostat. Recently, a self-shielded (Nieminen et al. 2011) pulsed superconducting coil (Vesanen et al. 2012a) has been demonstrated for ULF MRI; the B_p coil was integrated into the cryostat with the SQUIDs. The choice of materials for B_p can be important. For example, we have found that multi-stranded Litz wire performs much better than solid wire in terms of noise. According to Vesanen et al. (2012a), the superconducting wire magnetized if too high of a current (> 12 A) was applied, producing spurious gradients that influenced the image quality, limiting B_p in that work to < 24 mT.

Perhaps the most challenging aspect of a pulsed B_p is how to remove it appropriately. When the B_p field changes transient eddy currents will be induced in nearby conductors, which can impose a long dead-time if the magnetic fields from the transients exceed the dynamic range of the SQUIDs.

There are two approaches to switching off B_p: adiabatic and non-adiabatic. In a non-adiabatic ramp-down, dB_p/dt >> γB_m² such that the magnetization is left aligned with the original direction of B_p . If B_m is orthogonal to B_p, precession will begin automatically. In principle, this approach can minimize the time between beginning precession and measurement. In reality, however, the faster the ramp-down of B_p, the larger are the transients that are induced in nearby conductors. When measurements are made inside a conductive MSR, these transients can become a serious confound as they may have components that persist for hundreds of milliseconds (Vesanen et al. 2012b) and are inherently low frequency and hard to be removed from the MEG. Even in the absence of an MSR, anything conducting nearby will also support transients that may impact the image and/or impose long wait times. One added consideration in the non-adiabatic field removal is the non-uniformity of B_p. Not requiring a uniform B_p greatly simplifies the magnet design, but signal is lost because of this non-uniformity; when precession starts, the spins are not all in phase. In addition, there are technical problems associated with the requirement to dissipate the energy stored in the B_p coil.

If, instead, an adiabatic ramp (dB_p/dt << γB_m²) is used, the final magnetization will be aligned with low B_m field, which is easy to make uniform. Further, phase coherence is typically improved due to lack of transients. A spin-flip pulse is then required to start precession. In an adiabatic ramp, dB/dt is lower and thus there is less danger of heating metallic implants or affecting therapeutic electronic devices the subject may be carrying This is a special consideration when imaging near metal. In an ULF MRI system, B_p is typically 10 – 200 mT and it is removed within 10 – 100 ms (depending on the approach). Thus, the field change dB/dt may range from 0.1 – 20 T/s. Even with relatively high pre-polarization, the dB/dt in a ULF MRI system is typically lower than that in HF MRI systems. However, the adiabatic ramp-down takes longer, and signal is lost due to T₁ relaxation during that time.

The generation of the other magnetic fields required for imaging (B_m and gradients) is relatively simple given their low field strengths. Typically, simple wire-wound coils can be utilized. An interesting and important aspect regarding the measurement and gradient coils is how we power them. Because the noise level should be as low as possible (~ 1 fT/Hz is a typical goal) at the frequencies of interest, the current noise of the power supplies must also be as low as possible. While this sort of noise is not an issue at the frequencies of HF MRI, at ULF it can be a problem. Typically this has been dealt with by the use of batteries and heavily filtered circuits. Unfortunately, these solutions limit the sorts of pulse sequences that can be used. Thus, it is likely that one critical advance to ULF MRI instrumentation will be developing low-noise electronics that enable flexible pulse sequences. For example, the B_m field could also be oriented in any direction on the fly (enabling a totally new kind of projection imaging that would provide maximum sensitivity in a helmet-like configuration of sensors) but this approach also requires three B_m field coil sets to be on simultaneously.

Medical imaging

MRI at ultra low fields shows improved T1 contrast compared to high fields, which may translate in to a unique capability of ULF MRI to help delineate certain tissue types better than high-field MRI. It has already been shown that, for example, biopsies of prostate cancer tissue exhibit a significantly shorter T1 time at B_m = 132 µT than healthy tissue (Clarke et al. 2007) and cancerous vs. normal rat liver shows similarly high contrast at B_m = 100 µT (Liao et al. 2010). The complex behavior of the proton relaxation dispersion of water when approaching zero field (Hartwig et al. 2011) may be the physical background of the enhancement of T₁ contrast between healthy tissue and tumors at low fields.

The grey white matter border has a relatively high T₁ contrast. This border is blurred at focal cortical dysplasias (FCDs), malformations generated during cortical development. FCDs are highly epileptogenic and frequently cause intractable epilepsy. Unfortunately, a considerable fraction of FCDs can not be discerned in high-field MRI and are only detected in tissue microscopy after removal of the cortical region. Thanks to its higher T₁ contrast, ULF MRI may be able to visualize FCDs better than high-field MRI.

High-field MRI cannot be applied to patients with pacemakers, stimulators or metal in the body. However, there are no such restrictions with ULF MRI, which is inherently safe.

Temperature mapping

Vesanen et al. (2012) utilized the dependence of T₁ relaxation time of agarose gel to demonstrate the ability of ULF MRI to measure temperature. Although this method can prove useful in special cases, one must bear in mind that in human tissue, T₁ is influenced much more by the detailed structure of the tissue than by temperature. Therefore, this method can be used only for indicative purposes and to monitor possible changes of temperature when other factors affecting T₁ can be assumed fixed.

Conductivity imaging

If electric current is applied to a conducting object using two electrodes, the current will distribute itself between different paths in proportion to the conductivities of the paths. It was shown by Nieminen et al. (2012) and Vesanen et al. (2013) that the direction and amplitude of the current distribution can be measured with low-field MRI, provided, of course, that the signal-to-noise ratio is sufficient. This is not possible with a high-field MRI device without the need to turn the head between two measurements. So far, only simulation studies have been done. These indicate excellent reconstruction accuracy but also the fact that the signal-to-noise ratio of MRI measurements must be significantly improved before electrical impedance tomography can be done in humans with safe current strengths (on the order of 2 mA as those in transcranial direct current stimulation or tDCS).

Improvements in instrumentation

The present state of the art is not yet sufficient for scientific and clincal applications of combined MEG and ULF MRI. However, we can predict that ULF MRI can be improved to a level that will provide acceptable image quality and allow accurate registration of MEG and MRI coordinate systems. A great improvement in signal-to-noise ratio can be obtained by increasing the pre-polarization field strength and by reducing SQUID and dewar noise further; B_p > 100 mT and sensor noise level of 0.5 fT/Hz^1/2 seem possible even in a large array. The improvement in signal-to-noise ratio may enable us to measure the conductivity structure of the head as well. However, problems arising from the highly sensitive SQUIDs in strong pulsed magnet fields will become more severe. Thus, the task of building a practical MEG-MRI system is far from trivial. Elaborate methods will be needed to handle problems caused by eddy currents in the system and in nearby structures as well as the magnetization of materials. Optimized sequences and signal processing will be needed to maximally utilize the recorded data. Next, we will give a glimpse of one approach in developing signal processing.

Advances in signal processing

Conceivably, ULF MRI with up to hundreds of sensors for whole-brain imaging will be developed in the next few years. In high-field MRI, such highly parallel signal detection has been used to improve the spatiotemporal resolution at the cost of SNR (Pruessmann et al. 1999; Sodickson and Manning 1997). The feasibility of 3-fold acceleration has been reported in an ULF MRI study with a seven-channel SQUID system (Zotev et al. 2008a). Although currently SNR of ULF MRI is too low to be compromized, we expect that with future SNR improvement offered by higher pre-polarization field and more sensitive SQUID detectors, it might be possible to trade-off SNR for a shorter acquisition time. The SNR loss in parallel MRI (pMRI) is the consequence of the loss of data samples and the noise amplification in image reconstruction (Pruessmann et al. 1999). While the former loss is inevitable in acceleration, the latter loss can be compensated for by regularized reconstruction methods (Lin et al. 2004; 2005) and by increasing the number of parallel detectors.

The SNR penalty of accelerated ULF MRI was recently investigated by simulating a helmet-shaped sensor array with up to 306 SQUID sensors (204 gradiometers and 102 magnetometers; VectorViewTM, Elekta Oy, Helsinki, Finland) (Lin et al. 2013). It was found that, in all SQUID sensor geometries, image locations away from pick-up coils show a larger g-factor in general. These numbers suggest that, in a 3D ULF-MRI acquisition with two phase-encoding directions, the highest acceptable acceleration rate may be from 9- to 16-fold (based on an arbitrary threshold of average g = 1.4). Consider a 3D ULF-MRI acquisition with 64 x 64 x 64 voxels. It needs 4,096 independent read-outs with 64 x 64 phase encoding steps. With repetition time TR = 1 s, this amounts to more than an hour of acquisition time. Using an array of 102 sensors and 9-fold acceleration and assuming that the SNR loss due to reduced samples is tolerable, the data acquisition can be completed in approximately 7 minutes. However, as suggested by these results, spatially varying noise amplifications (i.e., g-factor) could be significant at 9- and 16-fold acceleration, resulting in inhomogeneous image quality deterioration.

In addition to or instead of aiming at achieving a higher resolution or shorter measurement time, parallel MRI can be applied to ULF MRI to improve the SNR by exploiting the redundancy among the receiver channels by enforcing k-space data consistency among them and by adding a priori image sparsity information to further suppress noise (Lin et al. 2013).

By performing MRI measurements with the large arrays of SQUID sensors available in MEG helmets, one can realize combined MEG and MRI, which offers unprecedented possibilities to obtain new kinds of information about the human brain. MEG-MRI systems will be quiet, open, and safe. They will enable highly accurate registration of MEG and MRI coordinate systems and, if imaging of injected current density proves practical, the determination of the three-dimensional conductivity distribution. This, in turn, would enable us to solve the inverse problem of MEG (and EEG) using reliable knowledge of both measurement and conductivity geometry. However, we are still far from constructing practical devices. It will be necessary to improve the signal-to-noise ratio considerably to attain MR image quality that is sufficient for clinical and scientific applications.