Magnetic Resonance Imaging technology – bridging the gap between noninvasive human imaging and optical microscopy

Abstract

Technological advances in Magnetic Resonance Imaging (MRI) have provided substantial gains in the sensitivity and specificity of functional neuroimaging. Mounting evidence demonstrates that the hemodynamic changes utilized in functional MRI can be far more spatially and thus neuronally specific than previously believed. This has motivated a push toward novel, high-resolution MR imaging strategies that can match this biological resolution limit while recording from the entire human brain. While sensitivity increases are a necessary component, new MR encoding technologies are required to convert improved sensitivity into higher resolution. These new sampling strategies improve image acquisition efficiency and enable increased image encoding in the time-frame needed to follow hemodynamic changes associated with brain activation.

INTRODUCTION

The traditional strength of functional MRI (fMRI) has been its ability to non-invasively image hemodynamic responses including changes in blood flow, volume and oxygenation associated with brain function over large regions, including the entire human brain. Although MRI technology has focused on methods allowing the study of smaller and smaller structures while retaining its whole-brain coverage, invasive methods using multi-photon and light-sheet fluorescence microscopy have an inherent ability to image down to the neuron scale but have expanded on the opposite front—moving from small fields of view to larger territories, including entire brain regions [1–5] (see Fig. 1).

Fig. 1. Scales of functional architecture accessible to fMRI in humans and to optical imaging in animal models.

Although continued technological development has decreased the spatial resolution of human fMRI to the submillimeter scale, further developments will be required to resolve units of cortical functional architecture such as cortical columns and layers over the entire brain. At the same time, invasive optical imaging techniques have expanded their spatial coverage to several millimeters, and in some cases an image plane covering a slice of the entire two cerebral hemispheres. Recent studies have shown that as these two complementary imaging modalities converge they can inform one another, yielding new insights into brain function. Further technological advances are required to continue to enhance the spatial resolution and neuronal specificity of human fMRI while retaining whole-brain coverage.

As these two technologies meet, they will eliminate a gap in our in vivo imaging capability, providing an unprecedented opportunity to translate insights from small animal studies directly to human studies and also greatly inform the mechanisms underpinning fMRI. For example, functional connectivity derived from resting-state fMRI has withstood many indirect validations, but its fundamental tenant—that correlations in the hemodynamic signals reflect correlated neuronal firing patterns—was only recently directly demonstrated using wide-field microscopy experiments [2]. This bridging of the scales is a powerful motivator to extend human neuroimaging with MRI toward studies of smaller-scale functional circuitry. The goal of this review is to examine recent MRI advances in MRI system hardware, data acquisition, and image reconstruction strategies that are being pursued to achieve this goal.

The push toward higher spatio-temporal resolution MRI is not new [6]; but it has not been straightforward to “turn up the magnification” to enhance the resolution of an MRI experiment. It is not only hardware limitations (i.e., a lack of bigger magnets) standing in the way, but also performance restrictions. Improved sampling strategies are needed to acquire the additional information at the highest spatio-temporal resolution allowed by hemodynamics. Increasing performance also requires attention to safe operation of the scanner for human subjects. In practice, these goals require coordinated improvements in all aspects of the MRI acquisition.

Current state-of-the-practice and potential biological limits of the spatial-temporal resolution of hemodynamics-based fMRI

Current practice (2–4 mm resolution in 3 Tesla magnets and 1–2 mm in 7 Tesla magnets) is firmly limited by MRI technology rather than any biological limit. In this review, we consider only isotropic spatial encoding strategies; in our view, isotropic voxels are needed to provide spatially-unbiased sampling over large brain regions. Invasive animal-model studies suggest that blood flow regulation is more precise both spatially [7,8] and temporally [9,10] than the 2–5 mm that have traditionally been considered the “hemodynamic limit” [11]. Hemodynamic control has been shown to be regulated at the 200 μm scale in the rodent olfactory bulb [12,13], and, in less-specialized cortex, blood flow regulation has been seen at the capillary or precapillary levels [14,15]. Importantly, these observations place the unit of regulatory control at a smaller scale than even current high-resolution fMRI studies, although an observation in one structure does not necessarily generalize to all brain regions. Nonetheless, these studies suggest that if improved instrumentation can be built, it will enable mapping of finer levels of neural activation.

Although the Blood-Oxygenation-Level-Dependent (BOLD) contrast is by far the most robust and commonly used signal used in fMRI, alternative techniques may provide enhanced neuronal specificity. Notably, fMRI based on changes in cerebral blood volume may increase sensitivity to the capillary bed, improving localization of neuronal activity [16,17]. This technology has been recently extended to human imaging [18–22]. New techniques aimed at non-hemodynamic signals such as the detection of cell swelling [23–25] or neuronal current oscillations [26–29] also hold promise. However, these new techniques require further investigation, are limited by low sensitivity, and face the same image resolution and coverage challenges as conventional fMRI.

INCREASING SENSITIVITY: BIG MAGNETS AND MASSIVELY-PARALLEL DETECTORS

The sensitivity of the detected MR signal decreases as the voxel volume; thus, a 3-mm cubic voxel has 27-fold higher signal level than a 1-mm cubic voxel. This rapid loss of signal amplitude immediately motivates improvements in the key technologies that define signal sensitivity: the static magnetic field (B₀) and the detection array coil.

Ultra-high magnetic field strengths

Not only does image signal-to-noise ratio (SNR) increase supra-linearly with magnetic field strength [30], but in the case of the standard fMRI signal—the BOLD contrast—the image contrast between the active and baseline states also increases, providing a substantial boost in contrast-to-noise ratio at high fields. Furthermore, the BOLD fMRI signal becomes more specific to the smallest intracortical blood vessels and capillaries with increasing field strength [31]. Thus, at high fields, BOLD reflects signal changes in the vascular compartment where the coupling between neuronal activity and the hemodynamic signals is tightest, therefore the neuronal specificity of fMRI is enhanced at high fields. These are the key motivations for the recent push for ultra-high-field MRI—defined as MRI at fields above 7 Tesla—for functional neuroimaging.

Increasing the magnetic field is the most direct means to increase sensitivity and specificity, but with higher fields come many challenges for imaging, including: artifacts and distortions associated with magnetic susceptibility and transmission wavelength effects in the human head, safety limitations due to radiofrequency (RF) power deposition, and amplification of physiological noise. Fortunately, concerted engineering efforts have addressed many of these issues, several of which are summarized below and in recent review articles [32]. An example of the improvements in image quality afforded by these efforts is presented in Fig. 2. Because of the success of the current generation of 7 Tesla and 9.4 Tesla human MRI scanners, recent efforts have evaluated the safety and engineering challenges of even higher magnetic field strengths (10.5 Tesla and 11.7 Tesla), and conclude that fields up to 20 Tesla would be feasible and would provide even further gains in sensitivity and specificity, opening the door to new applications [33–35].

Fig. 2. Recent progress in fMRI spatial resolution and image quality.

(Left) A standard 3 Tesla 3.1 mm isotropic echo-planar image used for conventional fMRI studies circa 2005. (Right) A more recent example of 7 Tesla 0.75 mm isotropic echo-planar image, acquired with parallel imaging acceleration and a 32-channel RF coil array and a specialized head gradient coil to achieve a four-fold accelerated rapid readout. Continued technology development will be needed to surpass this resolution for whole-brain neuroimaging studies.

RF coil arrays

Perhaps the single most significant engineering solution to the challenges of ultra-high-field MRI is the advent of RF receive coil arrays comprised of many detector channels operating in parallel, which arrived just after the introduction of 7 Tesla human scanners. There are two advantages to using a large number of smaller detectors to cover the head: SNR is increased (especially in the cortex); and parallel imaging performance improves, facilitating fast image encoding. Combining these into a single metric, we consider accelerated SNR, or the SNR of the accelerated, undersampled acquisitions used in parallel imaging, which has been shown to increase rapidly with higher-channel-count arrays [36]. The acceleration of the fMRI acquisition reduces artifacts such as image blurring and distortion [37–39]. This has inspired a trend towards larger brain arrays [40–42] and also dense arrays tailored to specific brain regions, such as 32-channel coils centered on the visual cortex or motor cortex [43–45].

NEW SAMPLING STRATEGIES TO OVERCOME ENCODING LIMITS

Although parallel imaging techniques have allowed the exploitation of data undersampling to pack in more image encoding per unit time, new image reconstruction strategies emphasize the role of the undersampling pattern design to reduce inefficiencies or redundancies in the encoding. These ideas exploit older concepts about the efficiency of 3D image encoding, and very benign forms of prior knowledge (such as “the head is round-ish”) as well as more strict assumptions about image compressibility. The result is more encoding per unit time—a critical goal for bringing instrumental resolution down to the biological limits. Several of these undersampling strategies are highlighted in Fig. 3 and explained further below.

Fig. 3. Sampling schemes used for efficient image encoding in accelerated MRI.

The time taken to encode an image can be reduced by skipping samples in Fourier or k-space, and the aliasing artifacts induced by this undersampling can be removed during image reconstruction through parallel imaging techniques. This allows one to accelerate the acquisition and still (ideally) produce an alias-free image. The distribution of aliasing in the image however is determined by the sampling pattern in k-space, and some patterns of aliasing are more amenable to removal via parallel imaging reconstruction than others. (A) Example of the application of CAIPI to Simultaneous Multi-Slice EPI. In this example, three axial slices are acquired simultaneously; the locations of the three example slices are shown in red in the sagittal reformat of the data. During acquisition, “blipping” the z-gradient imposes a slice-specific shift such that the acquired, collapsed images exhibit less overlap compared to the case without CAIPI shifting. (Adapted from Ref. [62], with permission.) Acquiring multiple EPI slices simultaneously increases the image encoding efficiency and therefore increases temporal resolution. (B) Illustration of undersampling schemes applicable to 2D or 3D imaging. Shown are images directly reconstructed from the undersampled data to demonstrate the pattern of aliasing that must be removed in each case (top row). Simplified schematics of the corresponding undersampling patterns in k-space are provided (bottom row). In each case an undersampling factor of 2 is applied to 3D encoded data. In the case of CAIPI undersampling, the data are still undersampled by a factor of 2 but the samples are shifted into a “quincunx” pattern to cause aliased replicates to fall into the corners of the image. In the case of wave-CAIPI, the same “quincunx” undersampling is performed combined with “wave” readouts to spread aliasing into the readout direction (head-to-foot) as well; to demonstrate this effect, the insets shown in the original image (for reference) and in the wave-CAIPI image show a sagittal slice through the volume at the location of the dashed green line. In the irregular Compressed Sensing undersampling, the sampling pattern is pseudo-random but still achieves an overall undersampling factor of 2; in this case the aliasing appears incoherent or noise-like in the direct reconstruction.

Controlled Aliasing in Parallel Imaging (CAIPI) and Simultaneous Multi-Slice Imaging

Although early parallel imaging methods provided huge benefits to fMRI, Controlled Aliasing In Parallel Imaging (“CAIPI”), and the Simultaneous Multi-Slice methods it has enabled, are probably the most impactful second-generation parallel imaging methods. The goal of parallel imaging reconstruction is to remove aliasing incurred by undersampling the acquisition through leveraging the partially redundant information afforded by the coil array [46,47]. The pattern of the aliasing, however, is determined by the sampling pattern. In the CAIPI approach [48], the overall degree of image aliasing is lowered by forcing the aliasing replicates to fall in signal-void regions e.g. outside of the head, yielding a simpler reconstruction with fewer residual artifacts and less noise amplification. This concept has provided tremendous gains in anatomical imaging, especially in applications using 3D image encoding [49]. The sampling strategy can also be changed dynamically when combining data across multiple repeated acquisitions [50] or different image contrasts [51] to further suppress artifacts enabling even faster acquisitions.

The concept of CAIPI can be extended to the readout direction by utilizing “zig-zag” [52], “bunched phase encoding” [53], or “corkscrew”/“wave” [54–56] k-space sampling trajectories. This further spreads out aliasing in multiple directions to provide increased imaging performance, providing over an order of magnitude speed-up compared to conventional acquisitions [57,58].

Simultaneous Multi-Slice imaging [59,60] is a parallel imaging technique that can also benefit from CAIPI [48,61]. Simultaneous Multi-Slice imaging was recently extended to echo-planar imaging (EPI), the “snapshot” imaging method most commonly used for fMRI, and has already provided major benefits to fMRI [62–64]. The addition of CAIPI to Simultaneous Multi-Slice EPI was attempted early on [60] and was ultimately achieved by a balanced z-gradient blip train to impose the CAIPI slice shifts, providing dramatic improvements in image quality and SNR [62]. This has enabled acceleration factors of up to 10 simultaneous slices with conventional coil arrays [65]. Acceleration of encoding in the is an enabling technology for high isotropic resolution, whole-brain imaging. Without it, acquiring all the slices during the timescale afforded by hemodynamic changes (~3 s) is not possible.

Compressed Sensing

Compressed Sensing is a conceptually-related technique in which the sampling pattern is designed to be pseudo-random or irregular to impose unstructured aliasing (i.e., mimicking noise, see Fig. 3B). The pseudo-random sampling is combined with a prior assumption that the image can be represented in some transform domain with a sparse set of coefficients (i.e. the image information is highly compressible) [66]. Not only can Compressed Sensing techniques provide large gains in encoding efficiency for classes of images that can be sparsely represented in a transform domain (e.g., the wavelet domain or the Fourier domain), but their application to MRI merged the powerful concepts of tailored sampling and model-based reconstruction.

Magnetic Resonance Fingerprinting

Magnetic Resonance Fingerprinting is a recent concept that builds off of the insights of CAIPI, Compressed Sensing, and model-based reconstruction, with a goal of providing quantitative maps of tissue relaxation properties as potential imaging biomarkers [67]. In Magnetic Resonance Fingerprinting, the various pulse sequence parameters that control image contrast are varied in time pseudo-randomly. A model of the time-evolution of the signal as a function of the tissue properties of interest (e.g., longitudinal magnetization relaxation time T₁, irreversible transverse relaxation time T₂, apparent diffusion coefficient) and nuisance factors (e.g., local magnetic field offset ΔB₀, transmit field B₁⁻, coil sensitivity B₁⁺) and generates a “fingerprint”. A dictionary of pre-computed fingerprints is compared against the measured signal evolution of each voxel using fast pattern matching to find the tissue parameter values [67]. As with Compressed Sensing, Magnetic Resonance Fingerprinting exploits the incoherence of aliasing artifacts imparted by the pseudo-random acquisition—although any given image may have artifacts, the artifact pattern varies over time such that the final parameter maps have lower artifact levels. Although promising for clinical MRI, Magnetic Resonance Fingerprinting has not yet found a role in fMRI studies.

Image reconstruction with Machine Learning

Although the application of Machine Learning to MRI reconstruction is extremely new, recent demonstrations have shown that supervised Machine Learning techniques, such as the convolutional neural networks (which form the basis of many Deep Learning algorithms), can be adequately trained on example data and generalize sufficiently to reconstruct data not in the training set [68–71]. The key applications likely to benefit most or to be enabled by Machine Learning remain to be seen, but early indications suggest that Machine Learning may find a prominent role in MR image reconstruction, the design of efficient acquisition schemes, or in removing complex artifacts such as those caused by subject motion.

REDUCING ENCODING BURDEN

Although faster image encoding and tailored sampling schemes can help achieve higher spatiotemporal resolution while avoiding image artifacts and blurring, there are several technologies that can be applied to reducing the encoding burden, for example by either decreasing the required number of encoding steps by reducing the spatial extent of the image, or by homogenizing the magnetic field.

RF transmission techniques for reduced encoding

As discussed above, achieving high spatial resolution requires many encoding steps. Because MRI uses Fourier encoding, the full extent of the anatomy intersecting a slice must be encoded, which, for those cases where only a small brain region is under investigation, prolongs image encoding, reduces efficiency, and causes increased image blurring [32]. One common strategy to eliminate signal from tissue outside of the region of interest, and thereby reduce the required number of encoding steps, is to utilize “zoomed” imaging, which either selectively excites only the tissue within the region of interest, or selectively suppresses tissue outside of the region of interest [32]. This can dramatically shorten the encoding time to allow for higher resolution [72,73]. Emerging parallel transmission techniques can provide for sharper definition of these zoomed regions to reduce artifacts associated with imperfect selective excitation or signal suppression [74–76]. The additional degrees of freedom in parallel transmission can also be used to cancel the electric fields and thereby reduce Specific Absorption Rate and tissue heating associated with these RF pulses [77,78].

B₀ shimming

B₀ inhomogeneity places additional burden on image encoding by increasing the need for parallel imaging acceleration to reduce distortion and blurring. The field inhomogeneity arises from complex patterns of fields as tissue magnetize in the large applied magnetic field, especially those around air-tissue interfaces. The “shimming” process attempts to eliminate these field nonuniformities and homogenize the B₀ field. Although conventional B₀ shimming using spherical harmonic shim coils built into the scanner has been sufficient in the past, we now see the need for local arrays of shim coils placed close to the head. These provide additional degrees of freedom to efficiently homogenize the field [79,80]. A new “AC/DC” design has emerged that can integrate the shimming elements (“DC”) with the RF detector coils (“AC”) such that both share the same conductor and do not compete for precious real estate around the head [81,82]. Advances in high-field shimming and field control have recently been reviewed [83].

Magnetic field monitoring, corrections, and feedback

Image encoding requires accurate control of the magnetic fields within the head during readout, and any uncompensated field gradients—either from lack of full calibration of the gradient system to dynamic gradients created by subject respiration—can cause artifacts and modulations of the image time-series. Magnetic field sensing probes have been recently introduced to measure and correct for these imperfections [84,85]. These probes measure the field at a single point in space concurrently with the imaging, and arrays of probes can be used as a “field camera” for scanner calibration [86,87] or acquisition-specific corrections [88]. They can also be positioned around the subject during scanning to dynamically monitor the fields. This information can be used either to feed these signals back into the system for real-time correction [85] or to remove artifacts during the image reconstruction [89]. Recent demonstrations have shown that probes embedded within the patient table can measure the field changes induced by the subject’s cardiac cycle [90,91], which can be used to either prospectively synchronize the measurement to the heart rate to remove pulsatile blood flow or related motion artifacts, or retrospectively regress out physiological noise from the image data.

SPECIALIZED AND MULTI-MODAL MRI SCANNERS FOR NEUROSCIENCE

Since neuroscience (as opposed to clinical imaging) forms the sole use of a growing fleet of MRI scanners, there has been a recent trend towards specialized MRI scanners with hardware tailored to functional imaging for neuroscience studies.

Gradient coils for functional MRI: increasing slew rate

The performance of gradient coils was previously limited by engineering concerns, such as achieving a spatially-linear gradient field, removal of heat from the coil, and the efficiency and power requirements of the coil. These issues are now largely resolved, and the main limiting factor on whole-body gradient performance is the bioeffects of rapidly switching, large magnetic fields on the human body. Peripheral Nerve Stimulation (PNS) can be induced if the product of the gradient switching rate and strength generate electric fields large enough to trigger nerve firing, which manifests as discomfort or pain. PNS is therefore strictly regulated in MRI. A smaller head gradient coil that covers only the subject’s head and neck eliminates the induction of large current paths in the chest and back which are prone to stimulation and thus offers a simple strategy for reducing PNS in brain imaging. Although head gradient coils with fast switching speed and strong gradients have been in use for some time [92], the new sensitivity afforded by higher field strengths and the increasing need for faster encoding has created a renaissance in head gradient coil use [93–95] for fMRI.

Although gradient coil manufacturers have heuristics to design coils that avoid PNS, ultimately the ability of a coil to prevent PNS is only known once the coil is constructed and operated. Recent work has shown that numerical body models with explicit human nerve anatomy combined with nerve dynamics models can closely predict PNS levels measured in volunteers [96]. This approach has the potential to accelerate testing and save cost, by providing PNS predictions that can be fed back into the design process to potentially improve gradient coil performance applicable to human imaging.

Similar to RF coil arrays, “matrix gradient coils” provide additional spatial degrees of freedom for encoding, potentially allowing increased efficiency or PNS control [97,98]. Often these gradient coils relax the linearity constraint of the gradient field, introducing non-bijective, non-Fourier encoding [99,100]. Early attempts to utilize these novel encoding schemes are being explored in the PArallel imaging Technique using LOCalized gradients or “PatLoc” framework [101] and in O-space imaging [102].

Gradient coils for diffusion MRI: increasing gradient strength

For diffusion imaging, gradient coils capable of producing gradient fields with high amplitudes provide high degrees of diffusion encoding in less time. This translates into a large boost in signal level by reducing signal decay from T₂ relaxation. Thus diffusion has been a primary motivation for recent dedicated gradient coils with large maximum gradient amplitudes [103–107]. For example, the “Connectom” gradient coil (designed for the Human Connectome Project) utilized a segmented design with four power amplifiers per physical axis (so 12 total) to overcome efficiency limits [103]. Unlike high-resolution EPI readouts, diffusion encoding has relatively low slew-rate requirements, and therefore avoids PNS issues.

Combined modalities: PET-fMRI, EEG-fMRI, TMS-fMRI, and potential of MR-tFUS

Because the various human neuroimaging and neuromodulation modalities can access different aspects of brain function, ongoing work seeks to merge their unique capabilities to extract new information about human brain function and organization. Continued development of instrumentation and technologies for integrating fMRI with Positron Emission Tomography (PET) [108–112], Electro-Encephalography (EEG) [113–116], and Transcranial Magnetic Stimulation (TMS) [117,118] has addressed many of the challenges faced when introducing additional equipment into the MRI environment. Recent work combining MRI with ultrasound may pave the way to integration with transcranial Focused Ultrasound (tFUS) [119], and progress using ultrasound to stimulate neuron populations [120,121] may in the future provide key insights into brain function and neurovascular coupling once the mechanisms of neuronal stimulation are better understood [122,123]. These multi-modal approaches can be used to synergistically merge the hemodynamic signals measured in fMRI with complementary metabolic and electrical signals, and image whole-brain responses to targeted neuromodulation, to both enhance neuronal specificity and map large-scale brain circuitry.

CONCLUSIONS

Several themes emerge from reviewing recent trends and advancements in MRI technologies for neuroscience. MRI technology has been effective at utilizing increasing degrees of freedom both in the sensors (RF detectors, motion cameras, field probes) and effectors (RF transmitters, B₀ shim arrays, gradient arrays). There is also movement toward dedicated MRI scanners that excel at a desired form of imaging such as fMRI or MR diffusion imaging. New insights into increased encoding efficiency in image sampling have provided improved encoding and image reconstruction algorithms. Overall, these techniques have improved the sensitivity, resolution, and functional contrast to advance toward improved whole-brain, high-resolution non-invasive measurement of brain function in humans which maximally leverages the neuronal specificity of the fMRI signals.

HIGHLIGHTS.

• Spatial scale of hemodynamic response is finer than fMRI resolution achievable today

• Improving fMRI resolution requires increases in sensitivity and image encoding

• Sensitivity can be increased with higher magnetic fields and dense detector arrays

• Image encoding can be sped up with fast gradient coils & efficient sampling schemes

• These advances will narrow the gap between noninvasive fMRI and optical imaging