High-resolution fMRI at 7 Tesla: challenges, promises and recent developments for individual-focused fMRI studies

Abstract

Limited detection power has been a bottleneck for subject-specific functional MRI (fMRI) studies, however the higher signal-to-noise ratio afforded by ultra-high magnetic fields (≥ 7 Tesla) provides levels of sensitivity and resolution needed to study individual subjects. What may be surprising is that higher imaging resolution may provide both higher specificity and sensitivity due to reductions in partial volume effects and reduced physiological noise. However, challenges remain to ensure high data quality and to reduce variability in ultra-high field fMRI. We discuss session-specific biases including those caused by factors related to instrumentation, anatomy, and physiology—which can translate into variability across sessions—and how to minimize these to help ultra-high field fMRI reach its full potential for individual-focused studies.

Introduction

In magnetic resonance imaging (MRI), image signal-to-noise ratio (SNR) increases with increasing magnetic field strength [1]. This higher sensitivity allows for acquisition strategies that enable higher image quality and resolution [2]. Additionally, higher field strength enhances magnetic susceptibility-based contrast, such as the blood-oxygenation-level dependent (BOLD) contrast—the most common contrast used for functional MRI (fMRI). The boost in functional contrast-to-noise ratio (fCNR) in BOLD has made fMRI a key application of ultra-high field (UHF) MRI at field strengths of 7 Tesla (7T) and above [3]. This sensitivity boost enables studies of fine-scale functional architecture, such as cortical columns, layers, and subcortical nuclei [4–6] and also increases statistical power needed for single-subject analysis. As discussed elsewhere in this Special Issue, a central motivation behind group studies in neuroimaging is to increase sensitivity to detect subtle functional features by averaging across subjects. However, meaningful individual differences, in structure or functional organization, can smear out these features when averaging across subjects. Thus the increased statistical power of BOLD-fMRI at UHF may also help enable the detection of subject-specific activation patterns in individuals.

UHF-fMRI, however, has its challenges. Several well-known fMRI artifacts increase in severity with field strength, and several subtle sources of bias are amplified—which, if unaddressed, can reduce sensitivity and counteract the gains of moving to higher fields. So, do the benefits of UHF-fMRI outweigh the costs? To address this question, here we review new technologies that improve UHF-fMRI quality and survey sources of bias. Notably, biases and artifacts can vary both within and across experimental sessions and thereby contribute variance or ‘noise’. We discuss how these sources of variance can be identified, characterized, and minimized. If these challenges are addressed, UHF-fMRI can be a powerful tool for individual-focused studies.

Benefits of 7T fMRI for individual-focused fMRI studies

The motivation behind individual-focused fMRI at UHF can be appreciated by reviewing the studies performed to date. In a pair of influential studies, first conducted at 3T then repeated at 7T, whole-brain activation to a simple visual stimulus was measured after averaging 100 runs across 10 experimental sessions [7,8]. Figure 1A summarizes how the measured activation pattern increasingly extended across the entire brain as more runs were included, well beyond the expected visual cortical and subcortical regions. The large number of samples allowed them to use trial-locked averaging, and this “unconstrained” analysis uncovered meaningful activations that are missed by conventional model-based analysis. The lesson was that meaningful activations that do not match our models may go undetected, leading to what they referred to as a pervasiveness of false negatives in standard studies. At 7T this analysis revealed activations that were more focal and contained within the cortical gray matter compared to 3T (see Fig. 1B).

Figure 1: Examples of novel insights from a highly-powered single-subject 7T fMRI data.

(A, B) Gonzalez-Castillo et al. show how time-locked model-free analysis of BOLD-fMRI data reveal whole-brain activation. The average task-related activation pattern extends beyond the visual areas with an increasing number of runs. The 7T activation pattern is more confined to the grey matter compared to 3T. (Adapted from [8].) (C) Braga et al. reveal how the Default Network fractionates into two spatially separated sub-networks within the individual using 31 sessions of resting-state BOLD-fMRI at 3T. (Adapted from [12].) (D) They repeated their study at 7T where they can reveal the network separation within the individual from a single session of 24 minutes worth of data. (Adapted from [12].)

Similarly, a well-known pair of highly-powered studies of individual subjects demonstrated that consistent patterns of functional connectivity were only achievable by pooling over 100 minutes of resting-state fMRI data acquired over many experimental sessions at 3T [9,10]. These studies revealed idiosyncratic functional connectivity patterns in individuals that were not seen at the group level. They concluded that between-session variability is dwarfed by between-subject variability, and that this inter-individual variability is a dominant confound in group-level differences. We return to the topic of between-session variability further below.

Another pair of highly-powered studies of individual subjects revealed subject-specific patterns of subnetworks that challenge widely-held theories of brain network organization. Again first conducted at 3T over 24 experimental sessions [11] and then replicated at 7T [12], the Default Network (DN) was found to fractionate into two independent spatially interdigitated sub-networks, shown in Figure 1C. The broad anatomical location of the networks was conserved across subjects, but the subnetwork spatial pattern details showed features idiosyncratic to the subject that were consistent across sessions. These subnetworks disappeared when the data were pooled across subjects [11]. Remarkably, these same subnetworks were observable in the individual from a single experimental session at 7T [12], showcasing the ability of 7T to provide sufficient sensitivity for single-subject fMRI. Figure 1D shows the subnetworks of an individual derived from 7T data, where the fine-scale detail can be appreciated.

Individual-focused fMRI is an emerging direction for whole-brain fMRI studies, but it is standard for studies of meso-scale functional architecture such as cortical columns [13–16], as the precise spatial layouts of these patterns are expected to be unique to the individual and thus require single-subject analysis. Although these high-resolution fMRI studies are performed at UHF, because mapping columns often requires detection of subtle differences in functional activation, extensive averaging is still required. One recent study into the interdigitated columnar system in human visual cortex showcased the ability to resolve these fine-scale details only when sufficient BOLD-fMRI data from multiple sessions were carefully acquired and analyzed [16]. This is somewhat analogous to the abovementioned study revealing interdigitated subnetworks by carefully pooling data across sessions [11].

Moving from 3T to 7T increases sensitivity only if small fMRI voxels are used—and in fact one may experience a sensitivity loss at 7T if large fMRI voxels are used! This is because physiological noise, which is a structured noise fluctuation in the fMRI signal, also increases with field strength [17], diminishing sensitivity gains. However, small fMRI voxels have reduced physiological noise and are thermal-noise dominated, therefore high-resolution fMRI acquisitions enjoy the expected sensitivity boost at higher field [18,19]. The critical voxel-size cutoff for thermal noise dominance depends on several factors [19], however a rough estimate for an upper threshold for BOLD-fMRI is 1.5 mm isotropic, and studies using voxels larger than this threshold may not benefit when moving from 3T to 7T.

One may then ask whether conventional studies that do not require high resolution per se would benefit from moving to higher field. While general recommendations are difficult to make, reducing voxel size can also benefit these conventional studies. This is because smaller voxel sizes reduce partial volume effects, allowing the voxels to sample gray matter more exclusively. Perhaps surprisingly, small-voxel fMRI with advanced anatomically-informed smoothing can provide higher overall sensitivity and fCNR than conventional large-sized voxels [20]. Therefore, acquiring fMRI data with voxels small enough to adequately sample the cortex is advantageous even for conventional studies investigating large-scale brain organization.

High-performance EPI at 7T

The most widely used acquisition method for fMRI is echo-planar imaging (EPI). The main limitation of EPI is geometric distortion and blurring along the phase-encoding direction. This distortion predominantly arises from magnetic field inhomogeneities caused by tissue susceptibility gradients. The severity increases with field strength and can produce centimeters of artifactual tissue displacement in the brain at 7T. The most effective tool to minimize distortions is accelerated parallel imaging, which increases the encoding bandwidth in the phase-encoding direction. Parallel imaging performance increases with field strength [21] and modern Radio Frequency (RF) receive coil arrays [22,23] and acceleration factors up to R=4 are feasible. The distortion reduction with acceleration and improvement in data quality can be seen in Figure 2A.

Figure 2: Geometric distortion in EPI for different parallel acceleration factors and reconstruction strategies.

(A) The unaccelerated (R=1) image shows severe distortion in particular in the frontal area (red box) and the ear canals (yellow box). With increased acceleration the effective echo spacing decreases (which increases the bandwidth in the phase-encoding direction) leading to reduced vulnerability to distortion. (B) Advanced reconstruction methods for parallel imaging (here, Dual-Polarity GRAPPA or DPG) reduce ghosting artifacts that vary over time, thereby improving the temporal stability of the fMRI data. Reduced interference pattern artifacts are seen near to regions of B₀ inhomogeneity (white arrows) in DPG reconstruction, and the artifact changes slightly over time. Temporal SNR (tSNR) maps show higher sensitivity in DPG reconstruction in this same region after artifact removal. Below the ratio of the time-series instability of the standard and DPG reconstructions is shown, demonstrating improved temporal stability in the DPG reconstruction. (Adapted from [27].)

Acceleration also enables higher imaging resolution in EPI, thus parallel imaging is used ubiquitously in UHF-fMRI. Several recent pulse sequence and image reconstruction advances provide enhanced image quality and practical usability of high-resolution EPI, including improvements in motion robustness, reliability of highly-accelerated EPI [24–26] and reduction of subtle artifacts seen in UHF-fMRI [27,28]. Additionally these high-performance EPI methods can reduce within-run and across-run variability because these artifacts can vary over time [27,29] (see Fig. 2B). Implementing these technologies requires changes to the acquisition software, which requires specific expertise and resources, however several of these technologies listed above are now available on some UHF platforms, provided either by the vendor or freely distributed by research centers.

Another more overlooked source of variability is dynamic distortion in EPI. This is caused by an interaction between head motion and magnetic field inhomogeneity [30]. Motion causes distortion changes because the field offsets are a function of head orientation within the B₀ field [31]. Available post-processing correction techniques either estimate and correct the temporally-varying magnetic field patterns [32,33] or jointly estimate head motion and distortion [34]. Correcting dynamic distortions during the acquisition may be possible by combining dynamic field sensing [35,36], field control [37,38] and prospective motion correction [39,40]; several MR instrumentation companies offer these hardware-based solutions, however there are still challenges in integrating these into standard workflows for routine imaging. Motion suppression using custom head holders can also help minimize this challenging artifact [41,42].

The MRI scanner instrumentation places limits on the achievable imaging resolution and it is not always possible to acquire the desired spatial resolution, demanding new advances in MRI hardware (as recently reviewed [43]). Exciting innovations in UHF-fMRI instrumentation are ongoing [23,38,44], and are expected to have a major impact on imaging capabilities once they become widely available.

Bias from MRI instrumentation

Several sources of instrumentation bias become pronounced at higher field and cause spatially-varying fMRI sensitivity that induces between-session variability. The most well-known and perhaps strongest bias stems from the RF receive coil array. The receive sensitivity, or the B₁⁻ field, is highest closest to the detector and falls off steeply with distance [45] and thus detectability is maximal at the brain periphery and lowest at the center. Figure 3A shows the average B₁⁻ field in the standard 32-channel RF coil at 7T, with an 8-fold variation seen across the brain. This causes a position-dependent detection bias for high-resolution fMRI where the small voxels are thermal-noise dominated [46]. This also interacts with motion as the head moves through the spatially-varying coil sensitivity [47,48]. While correction techniques have been proposed [48], motion suppression may be the most effective solution. Inconsistent subject repositioning between sessions will cause sensitivity-induced variance; Figure 3B demonstrates how the B₁⁻ field varies up to 13% for a given subject between sessions. However this within-subject variability is lower than across-subject variability as shown in Figure 3C, indicating that the across-subject differences in brain anatomy/geometry and differences in head placement contribute more to B₁⁻ variability.

Figure 3: Example receive coil or B₁⁻ fields on within- and across-subject variability.

B₁⁻ field estimated for a standard 7T 32-channel coil array (derived from 20 subjects). (A) The B₁⁻ profile averaged across subjects shows a nearly 8-fold spatial variation in receive sensitivity, with folds closest to the periphery exhibiting consistently higher sensitivity. (B) The average difference in B₁⁻ profile between two sessions. The temporal lobes and folds in the periphery exhibit the highest difference in receive field sensitivity when the subject is repositioned in the scanner. The pattern of the B₁⁻ between-session difference is distinct from the average B₁⁻ profile, which likely reflects an interaction between the steep B₁⁻ profile of this coil array and the variability in head positioning, which itself will depend on the operator and on the coil array housing. (C) The B₁⁻ profile variability across subjects. The pattern of the variability of the B₁⁻ profile across subjects is similar to the between-session difference for the single subject, but 2 to 3 times more severe, likely because of the additional variability of brain geometry across subjects. This is particularly observable in the front of the brain which likely reflects the variation of the distance between forehead and the coil housing across subjects.

A related but distinct bias source stems from the RF transmit coil. With increasing field strength the RF transmit or B₁⁺ field becomes more spatially nonuniform due to dielectric effects. In practice this leads to a spatially varying flip angle, which can also cause a position-dependent fMRI detection bias in thermal-noise-dominated acquisitions [49]. It is impossible to achieve the desired flip angle over the entire brain, but one can either optimize the B₁⁺ field for a given region of interest (at the expense of others) or calibrate such that acceptable flip angles are achieved across the brain. This calibration procedure is vendor specific, and may require intervention by the scanner operator to ensure optimal B₁⁺ settings.

A final form of bias stems from gradient nonlinearity. MR image encoding assumes linear magnetic field gradients imposed over space by the gradient coils. Perfect linearity cannot be achieved in practice and there is spatially-varying nonlinearity in the gradient field. This causes geometric image distortion and spatially-varying voxel sizes. The gradient is linear close to the scanner isocenter but typically becomes nonlinear about 10 cm offcenter. Fortunately this gradient nonlinearity is determined by the gradient coil and is constant, therefore it is straightforward to account for in postprocessing [50]. Nevertheless, when comparing data across sessions, when the head may be in different positions relative to the isocenter, gradient nonlinearity correction will be required to remove differential distortion. Changes in head position during an experimental session can also cause differential distortion in brain areas near to the regions of nonlinear gradients.

All of these three sources of bias—RF receive field (B₁⁻), RF transmit field (B₁⁺) and gradient non-linearities—are impacted by head positioning, and therefore each will vary across experimental sessions, introducing between-session variability. While B₁⁻ and B₁⁺ fields can be estimated with relatively short acquisitions, and gradient nonlinearity fields are often provided by the manufacturer, these biases are not commonly addressed. One simple solution, analogous to motion suppression, is consistent head positioning across sessions, including positioning within the RF coil housing as well as landmarking of the subject and positioning of the patient table relative to the magnet isocenter. Custom-molded foam padding can help ensure consistent head placement, and has recently been used to minimize bias differences across sessions at 3T¹. Consistent operational procedures for scanner calibration and subject handling can also reduce variability, however more efforts are needed to provide analysis tools for characterizing and/or removing this variability across sessions.

Anatomical and physiological sources of bias

In addition to bias imposed by instrumentation, there are several subject-specific anatomical/physiological factors that are increasingly understood to impose spatially-varying bias on the fMRI signal. Partial volume effects vary across the brain as the tissue geometry changes relative to the voxel grid. For this reason, cortical thickness may impact BOLD contrast and noise [51]. Figure 4 shows the variation in the percentage of the fMRI voxel containing gray matter; the proportion and degree of gray matter voxels showing partial volume effects decreases with smaller voxels, reducing signal dilution from white matter and noise contamination from cerebrospinal fluid.

Figure 4: Spatially varying partial volume effects across the brain.

The percentage of gray matter contained within voxels intersecting the surface at mid-cortical depth is plotted across several isotropic voxel sizes, visualized on the inflated surface, with histograms displaying the distribution over the hemisphere. (Partial volume was computed using FreeSurfer for multiple fMRI voxel grids at different positions relative to the cortex, then averaged.) For the 3 mm case there are regions where fMRI voxels contain as little as 20% gray matter, while for the 1 mm case the fMRI voxels within nearly every region contain almost 100% gray matter. Thinner cortical regions (such as primary somatosensory cortex) exhibit lower percentages of gray matter. While the spatial heterogeneity of these effects is pronounced for larger voxels, partial volume can be readily estimated in order to take this effect into account during data analysis and interpretation.

Other spatially-varying biases include the cortical orientation effect, where the coupling between the vascular anatomy and cortical anatomy impart an influence on the BOLD contrast and noise that varies with cortical orientation [52,53]. This effect also increases with field strength [52]. Like the instrumentation biases mentioned above, the orientation effect changes with head positioning, posing an additional potential source of between-session variability even in individual-focused studies. Furthermore, vascular anatomy is known to vary substantially across the brain and across individuals [54].

Accounting for spatially-varying detection power

The biases discussed above produce spatially-varying statistical power that must be accounted for in both single-subject and group-level analysis. Consider a normal distribution of voxels having no effect N(0,1) (null hypothesis) and of voxels having a true effect N(√nμ/σ, 1)) where n is the number of samples, μ is the effect size and σ the noise standard deviation. Spatially-varying biases impart spatially-varying σ and a subsequent variation of statistical power across the brain. Bias-induced variations in σ are substantial, in particular at 7T. We recently demonstrated that cortical orientation effects in 7T fMRI data impart a 70% variation in fluctuation amplitudes across the brain [52], such that a cortical location with parallel surface orientation would thus require n_parallel/n_{perpendicular}=1.7² ≈ 3 times the samples to reach equivalent statistical significance compared to a location with perpendicular surface orientation. Biases also impart spatial variability in detectability between sessions and across subjects. Arguably, the between-session variability for an individual is lower than the variability between two subjects because brain geometry contributes to across-subject variability. This can be appreciated in Figures 3B–C, where between-session difference in the B₁⁻ field within-subjects is shown to be less than the across-subject variability. Thus, for a given effect size, on the basis of B₁⁺ variability alone a group-level study consisting of single experimental sessions compared to an individual-level study comprising multiple experimental sessions would require roughly n_{across-subject}/n_{within-subject} = (30%/13%)² ≈ 5 times total more samples to achieve the same statistical power across the entire brain.

This spatially-varying detection power may not be a problem for studies that are highly-powered, but proper power analysis during experimental design should factor in these effects. Ideally practitioners would have estimates of these biases to help with data interpretation, and to guard against incorrect inference when, e.g., an expected effect is seen in a region of the brain with high detection power while an expected effect is not seen in a region with low detection power.

While best-practice scanner operation strategies can reduce variance from both anatomy/physiology and instrumentation, complete elimination is unlikely, therefore more effort will be required to both account for this variability and to interpret individual-focused fMRI data. For example, analyses could incorporate bias maps to weight the fMRI data or derive confidence limits. Indeed, recent work demonstrated that fMRI reliability is enhanced when across-run variability is accounted for in the general linear model [55], and strategies with explicit measures of biases are expected to improve reliability.

Conclusions and Outlook

UHF-fMRI offers increased sensitivity to aid individual-focused studies, however several sources of variability are enhanced at UHF which give rise to variations in statistical power across the brain as well as variations across sessions and across subjects. Operational strategies exist for minimizing variations of sensitivity over time (i.e., across sessions) however variations in sensitivity over space (i.e., across the brain) will persist. If these sources of variability are addressed, single-subject UHF-fMRI can achieve higher measurement consistency than group-level studies. This will likely enhance statistical power per sample for single-subject studies: a recent study found that about 30 fMRI runs from two individual subjects at 7T, analyzed with best-practice methods, produced comparable results to a 181-subject average [56], meaning in this case 15 times less data were required for the individual compared to the group. UHF-fMRI thus appears ready to provide more information about meaningful individual differences in functional organization [57].

Relevant UHF-fMRI topics that were not discussed here are the potential for higher spatial and temporal resolution [58–60]; increased microvascular specificity of BOLD-fMRI [61]; and the possibility of non-BOLD-fMRI techniques that promise higher neuronal specificity [62]. Also, preprocessing choices influence the final results [63] and resolution can be inadvertently lost during preprocessing [50,64]; these losses must be avoided to retain the detailed functional patterns seen in the individual.

Overall, UHF-fMRI is a mature tool whose enhanced sensitivity can contribute greatly to individual-focused neuroscience. It is our hope that the strategies outlined here for addressing the specific challenges of UHF-fMRI encourage future studies that exploit this powerful technology.