Comparison of BOLD and CBV using 3D EPI and 3D GRASE for cortical layer fMRI at 7T.

Abstract

Purpose:

Functional MRI (fMRI) at the mesoscale of cortical layers and columns requires both sensitivity and specificity, the latter of which can be compromised if the imaging method is affected by vascular artifacts, particularly cortical draining veins at the pial surface. Recent studies have shown that cerebral blood volume (CBV) imaging is more specific to the actual laminar locus of neural activity than BOLD imaging using standard gradient-echo (GE) echo planar imaging (EPI) sequences. Gradient and Spin Echo (GRASE) BOLD imaging has also shown greater specificity when compared with standard GE-EPI BOLD. Here we directly compare CBV and BOLD contrasts in high-resolution imaging of the primary motor cortex for laminar fMRI in four combinations of signal labeling, SS-SI-VASO (CBV) and BOLD, each with 3D GE-EPI and zoomed 3D-GRASE image readouts.

Methods:

Activations are measured using each sequence and contrast combination during a motor task. Activation profiles across cortical depth are measured to assess the sensitivity and specificity (pial bias) of each method.

Results:

Both CBV imaging using GE 3D-EPI and BOLD imaging using 3D-GRASE, show similar specificity and sensitivity and are thus useful tools for mesoscopic fMRI in the human cortex. The combination of GRASE and VASO did not demonstrate high levels of sensitivity, nor show increased specificity.

Conclusion:

3D-EPI with VASO contrast and 3D-GRASE with BOLD contrast both demonstrate sufficient sensitivity and specificity for laminar fMRI to be used by neuroscientists in a wide range of investigations of depth-dependent neural circuitry in the human brain.

Introduction

High-resolution imaging of neural architecture at the mesoscale is growing in importance, especially due to an increasing focus on cortical laminar analysis for understanding neuronal circuitry. Blood oxygenation level-dependent (BOLD) contrast using gradient echo (GE) echo planar imaging (EPI) is the most commonly used functional magnetic resonance imaging (fMRI) method to investigate neural activity in the human brain (1), and the development of ultra-high field (7T and above) scanners has allowed robust imaging of neuronal organization at the mesoscale (2–4). Recent results using 3D Gradient And Spin Echo (3D-GRASE) imaging (5,6), where orthogonal excitation and refocusing pulses are used to image a smaller area of cortex with higher resolution and specificity, have further demonstrated the possibility of functional imaging at this scale (7–10). Alongside this, Cerebral Blood Volume (CBV) imaging using the VAscular Space Occupancy (VASO) technique (11,12) has demonstrated the ability to resolve activity from different cortical layers (13,14). In this paper, different combinations of image readout (EPI, GRASE) and contrast (BOLD, CBV) are compared for sensitivity (i.e. the amount of activation measured) and specificity (i.e. how well the activation can be localized) in laminar fMRI.

BOLD contrast is a combination of blood T₂* or T₂ changes between blood and surrounding tissues following neural activation (1). The BOLD signal arises from both small vessels specific to local neural activity, and draining veins that are not locally specific (15). In both small and large vessels, the BOLD signal consists of multiple components: intravascular (IV), arising from blood oxygenation induced susceptibility effects, and extravascular (EV), arising from static dephasing induced by field inhomogeneities and dynamic averaging around small vessels. Whilst the indirect nature of the BOLD response, being based as it is on changes in hemodynamics (16), prevents measuring the activity of individual neurons, it has the potential of measuring the activity of cohorts of neurons arranged at the mesoscale (0.5 – 1mm), namely neurons arranged perpendicularly (columns) and tangentially (layers) to the cortical surface. Cortical columns are groups of neurons arranged perpendicularly to the cortical surface that share a common preference for a stimulus (17,18), for example ocular dominance columns in V1 (4,19). In addition to this perpendicular organization, the cortical sheet is divided parallel to the cortical surface into roughly six layers, based on the properties and connections of the cells contained within each layer. The connections between these layers within a patch of cortex, and to other brain regions, are thought to define the processing performed in that region of cortex. In this model, bottom-up and top-down information flow can be seperated by which layer it arrives at (20). Being able to non-invasively measure neuronal activity at this scale would allow important insights into the fundamental circuits underlying cognition.

Increases in field strength and advances in hardware now routinely allow sufficient signal to image the brain at mesoscale resolution, and at this level it is the origins of the BOLD signal itself that define the sampling resolution for fMRI. When performing fMRI at ultrahigh field (UHF, 7T and higher), the signal change from the EV component increases quadratically with field strength (21,22), while the IV signal in both small and large vessels significantly decreases because the T₂* or T₂ of blood is very short (21,23), leading to some increase in specificity of signal to the site of neuronal activity. However, with the most commonly used GE-EPI acquisitions at ultrahigh fields, the EV BOLD contrast from large vessels will still be present in the images. This is seen most commonly as bias towards the outer (pial) surface of the neocortex when measuring signal change at different cortical depths (3,10). Whilst careful experiment design, analysis and post-processing can recover some of the required specificity from GE BOLD acquisitions for laminar (2,3,24,25) or columnar (4,19,26,27) imaging, the specificity of the GE signal has not allowed robust mesoscale mapping to reveal neuronal circuitry in cortical columnar or laminar organization.

Using Spin Echo (SE) EPI for fMRI, a 180° refocusing pulse partially recovers signal loss from static field dephasing in the EV signal around large vessels (22,23,28). This reduces the venous contribution to BOLD, which increases the spatial specificity (22,23), but at the expense of overall lower sensitivity. Combining SE imaging with a high B_0, which offsets the loss in sensitivity via a general gain in SNR_, results in suppression of both the EV and IV components in large vessels (22,23,29), leading to further increases in specificity. SE-EPI acquisitions have been used successfully for fMRI at 7T for imaging layers in animals (16,30,31) and columns in humans (4,32). When using an EPI readout for SE imaging, high-resolution imaging with an extended field of view (FOV) in the phase encoding direction requires very long readout times, which introduces increasing T₂* weighting at the outer regions of k-space, reducing signal specificity. Segmented (4,32) or accelerated (7) imaging can shorten echo trains to mitigate this effect, at the expense of lower SNR due to increased noise effects. An alternative and complimentary approach is inner volume (zoomed) imaging, where a smaller FOV (33) is acquired to limit the length of the overall echo train. This is achieved in the SE-EPI pulse sequence by orthogonalizing the excitation and refocusing pulse planes, reducing FOV to a limited area of cortex (4,32) and allowing high-resolution imaging (albeit for a single slice).

GRASE (5,6) combines multiple SE refocusing pulses of a Carr-Purcell-Meiboom-Gill (CPMG) sequence with intervening EPI echo trains in single shot imaging. Sub-millimeter resolution fMRI can be achieved by combining inner volume zooming with single-shot 3D-GRASE (5) to reduce the FOV and shorten the multiple EPI echo trains to minimize the T₂* weighting (7). Zoomed 3D-GRASE has been used for both laminar (8,10,34,35) and columnar specific imaging (9,34,36). The multiple short echo trains of 3D-GRASE increase T₂ weighting and decrease T₂* weighting in outer k-space. Another difference from SE-EPI is that 3D-GRASE is a variant of the CPMG sequence and therefore can contribute stimulated echoes (STE) with T₁ contrast to BOLD mechanisms (37).

Cerebral blood volume (CBV) contrast in fMRI has been shown to have an advantage over BOLD by excluding venous signal contributions, and has been shown to be tightly coupled to metabolic changes associated with neural activity (38–40). Non-invasive CBV-weighted fMRI in humans can be achieved using the Vascular Space Occupancy (VASO) method (11,12), which measures changes in CBV by acquiring an image around the short period while blood signal is nulled after an inversion pulse, so that changes in blood volume lead to a proportional decrease in signal. Recently the technique has been used successfully at UHF (41) for laminar specific applications (13,14). 3D-GRASE can also be used as the readout sequence in CBV imaging, as used at low resolutions for fMRI at 3T (42–44) and in resting state experiments (45,46). However, 3D-GRASE VASO has not been developed or evaluated for mesoscale high resolution CBV-fMRI. Given that VASO imaging with EPI, and BOLD imaging with 3D-GRASE at 7T have both demonstrated the ability to resolve activity at the mesoscale, combining the two methods could be beneficial for high-resolution functional imaging.

In this work, we compare 3D GE-EPI and zoomed 3D-GRASE readouts for high-resolution laminar fMRI, using BOLD and VASO contrasts. For both VASO sequences, an interleaved, pair-wise acquisition sequence (Slice-selective slab inversion (SS-SI) VASO), as previously utilized for 7T laminar VASO imaging (13,14), was used. This method produces two fMRI time-series data sets acquired simultaneously: a blood nulled VASO time-series and a standard BOLD time-series. The second (not blood nulled) image is used to remove the BOLD contrast that exists in the first (blood-nulled VASO) image, a particular issue at high-field (11). We hypothesize that the T₂-weighted GRASE BOLD images would show layer profiles similar to the VASO images, with both having less pial surface bias than T₂*-weighted EPI BOLD. With regard to the VASO contrast, GRASE images have a single excitation at the exact time of blood nulling, meaning that all acquired k-space partitions will have identical CBV weighting (42–44). This is a potential advantage over VASO using EPI, where multiple excitations occur with varying degrees of separation from the blood nulling time, leading to a number of potential issues, such as blurring across the image (14,47). However, blurring across the image is also an issue for GRASE owing to T₂ relaxation during the readout, which could potentially limit the ability to resolve mesoscale activity (7,48).

Throughout the manuscript, 3D GE-EPI will be referred to as EPI, and zoomed 3D-GRASE will be referred to as GRASE. In addition, the following terms are used for four different image/contrast types:

• EPI-VASO and GRASE-VASO - images acquired with 3D GE-EPI and 3D-GRASE sequences, respectively, at the first inversion time (TI). At this time, blood signal is nulled to detect CBV. BOLD contribution has been corrected using the image from the second TI.

• EPI-BOLD and GRASE-BOLD - images acquired with EPI and GRASE sequences at the second TI, when blood signal is not nulled.

Methods

Data Acquisition

We analyzed activation-related signal change in the region of motor cortex representing index finger and thumb motion, and activation-related signal changes as a function of cortical depth, for all four combinations of contrast and acquisition (i.e., EPI-BOLD, EPI-VASO, GRASE-BOLD and GRASE-VASO). The study protocol was approved by the Institutional Review Board at the San Francisco VA Center; each participant gave written informed consent before MRI data acquisition. 6 healthy volunteers (age: 43.2 ± 13.7, 3 female) participated in this study, with 5 healthy volunteers (age: 38.3 ± 2.1, 1 female) participating in a second follow-up session.

Imaging Hardware

FMRI data was acquired on a MAGNETOM 7T scanner (Siemens Healthineers, Erlangen, Germany), with an SC72 body gradient coil (Gmax = 70mT/m and SR= 200 mT/m/ms, effectively set to Gmax = 42mT/m and SR= 190 mT/m/ms by inbuilt scanner software limits). RF reception and transmission were performed with a 1 Channel Transmit/32 Channel Receive Head Coil (Nova Medical, Wilmington, MA, USA).

Stimulation paradigm

To induce VASO and BOLD functional signal changes, a unilateral finger tapping (thumb and index finger) task paradigm, previously used to investigate layer specific activation (14), was utilized. In brief, it consisted of 12 blocks, each of 60 s duration (30 s tapping, 30 s resting), resulting in acquisition time (TA) of 12 min. Subjects were cued to tap their forefinger and thumb with the same pacing as a video animation (49) projected in the scanner bore. This paradigm has been shown to increase activity in Layer II/III (cortical input) and Layer Va (spinal output) in the hand-knob region in M1 (14). This expected pattern of activation can be used to assess the sensitivity, in this case the strength of activation seen in the hand-knob, and the specificity, the inverse degree of weighting towards the pial surface, of the different sequences used.

Pulse Sequence

A slice-selective slab-inversion (SS-SI) VASO sequence (11) with 3D-GRASE (6) readout was implemented. Specifically, the GRASE pulse sequence was modified to include an inversion recovery (IR) pulse to acquire an image with VASO contrast, followed by an additional readout to acquire a second image with BOLD contrast (Figure 1). The SS-SI VASO EPI sequence was implemented as per previous studies (14,47) with a 3D GE-EPI readout (50). In VASO, the signal decreases with an increase in blood volume, so CBV change is inversely proportional to the VASO signal change, and is expressed as CBV change in ml per 100ml of tissue (13,51). In the current experiment results were expressed in units of percent signal change to allow easier comparison with the BOLD results, across multiple sequences and with previous work (13,14,47).

Figure 1:

GRASE pulse sequence with two readouts (one with blood nulled, one without nulling) per TR, which are used to provide VASO and BOLD functional contrast. Inversion recovery pulses are drawn with red and GRASE image readouts are drawn in gray. Blood signal (red curve) is assumed to not be in steady-state. The signals of grey matter, white matter, and cerebrospinal fluid (blue, turquoise and black curves) are assumed to be in steady state after previous inversions and readouts, whereas fresh blood is inverted every TR. At the time of the image acquisition at TI1, blood signal is nulled, whereas at TI2, there is signal for all tissue types.

Comparison between EPI and GRASE readouts is shown in Figure 2. In EPI (Figure 2A), the 3D volume is excited using a small flip angle (FA) for each phase encoding step along the slice (partition) direction. The signal decays with T₂* along the in-plane phase encoding direction, so the final 3D image has T₂* contrast. The signal acquisition ordering in the partition direction was linear, with k₀ acquired at the time after the inversion pulse (inversion time, TI) at which blood signal was nulled, and the outer k-space partitions collected slightly before or after the blood nulling TI. In previous comparisons of linear and centric k-space ordering, with k₀ at the blood nulling TI, linear ordering was found to be more robust against temporal noise, to allow for shorter TRs by minimizing dead time, and to show less blurring compared to centric ordering (Figure S1 in (52)), hence was used here. In GRASE (Figure 2B), the 3D volume is excited only once and then the magnetization is refocused at each partition using nominal 180° RF pulses, with phase encoding in the partition direction. Along the partition direction the signal decays with T₂, and spin echoes are refocused at the k-space center line for each partition. Between refocused echoes (corresponding to the outer k-space phase encoding lines) the signal undergoes T₂* decay as well. Images acquired with GRASE readout have combined T₂* and T₂ contrast, leading to less bias towards large pial vessels.

Figure 2:

Comparison of two acquisition methods. A: 3D GE-EPI, B: 3D-GRASE. Note that the 3D-EPI image acquisition involves multiple excitation pulses (in blue) at slightly different Tis after the inversion pulse (IR), which can lead to variable T₁ weighting across the k-space partitions. 3D-GRASE has only a single excitation, leading to uniform T₁ weighting across k-space. GRASE greatly reduces T₂* in the image compared to EPI, given signal dephasing differences (T₂* Phase diff, simplified e.g. no effects of crusher gradients) are refocused multiple times to zero phase error, where spin echoes are created. Note that crusher and spoiling gradients are omitted from the figure for reasons of clarity.

With the inversion pulse necessary for VASO contrast, the EPI acquisition will have each k-space partition acquired at progressively later TIs, leading to different T₁ weightings across the slice-partition direction, which can result in blurring. This blurring can be mitigated by using variable FA (VFA) across the partition direction for each excitation (53), with the FA regime depending on the underlying tissue T₁, in this an assumed gray matter (GM) T₁ of 1800ms. This VFA method can only correct for the blurring in a single tissue component (GM) whereas white matter (WM) and Cerebral Spinal Fluid (CSF) have different T₁s. In addition, the VFA approach is limited by the inhomogeneities in RF transmit field (B1+) at 7T, hence the T₁-related blurring effect might be only partially corrected (14,47). GRASE signal, on the other hand, will be acquired at a single TI, established in time by the single excitation pulse, throughout the whole 3D k-space, so all partitions in the 3D volume will have exactly the same T₁ weighting. While blood signal will remain nulled (48) across the echo train, signal in other tissues will be affected by T₂ decay across the echo train, causing signal variations across slice partitions and hence blurring in the slice direction (7), which can be mitigated using VFA for the refocusing pulses at the cost of SNR (48). Excitation FA for the GRASE readout was set to 90°, while the refocusing FA had to be decreased to reduce power deposition and specific absorption rate (SAR) given the required 1.5s between image readouts, whilst maintaining a sharp refocusing slab profile. To optimally determine the refocusing flip angle, we investigated the effect of different constant refocusing flip angles on the SAR and SNR using the following metrics: 1) Relative SAR was estimated with increasing refocusing flip angles from 90° to 180° by calculating the sum of squares of the flip angles for the entire echo train relative to that with 180° (54):

$\text{rSAR}=\frac{1}{0.25+\text{ETL}}\left(0.25+{\displaystyle {\sum }_{\text{i}=1}^{\text{ETL}}{\left(\frac{{\text{α}}_{\text{i}}}{180}\right)}^2}\right)$

2) Relative SNR was calculated by integrating the modulation transfer function (MTF), divided by the SNR obtained with 180°. Supporting Information Figure S1A shows the behaviors of relative SNR and relative SAR for GRASE-VASO with increasing refocusing flip angles from 90° to 180°. It is observed that the relative SNR rapidly increases from 90° to 150°, reaching nearly maximum SNR level around 160° with a slow rate in a concave shape, while the relative SAR quadratically rises with increasing flip angles. Flip angles less than 180 will lead to STE contributions along the echo train in addition to the signal showing T₂ decay, which can recover some of the SNR lost with a smaller flip angle. Bearing this in mind, α_r = 165° was selected in the experimental studies by reducing the relative SAR with a slight loss of SNR, thus yielding the signal evolution fairly similar to that of α_r = 180° (Supporting Information Figure S1B) in terms of STE. Compared to a 180 ° pulse, using a lower flip angle can actually yield slightly improved blurring, although not to the extent of using a VFA acquisition (48). Deviations from the nominal flip angle (for example due to B1 inhomogeneities across the FOV) can lead to spatially dependent differences in blurring, temporal SNR (tSNR) and STE contributions for all GRASE acquisitions. This variation can actually have a beneficial effect on the blurring when using constant flip angles, while this is not the case for the lower flip angles used in VFA (48), or for the VFA used in the EPI VASO sequence (47). To assess the impact of VFA on a GRASE-VASO acquisition, the experiment was repeated on a second group of volunteers using a GRASE-VASO sequence with the same constant flip angle as the main experiment (165°) and one with VFA refocusing pulses (64–100°). Both CFA and VFA GRASE will generated STE as well as SE, with the effect being larger for VFA (7,48).

To minimize the duration of the EPI readout and therefore the T₂* decay, while achieving high spatial resolution, an inner volume (zoomed) acquisition (33) was implemented for GRASE (Figure 3). The FOV for GRASE was 99×25×12 mm³ and matrix size was 132×34×8, yielding a nominal resolution of 0.75×0.75×1.5 mm³; TE was 48ms. Partial Fourier sampling (55) (with zero-filling of the remaining k-space) of 5/8 was used in partition direction to reduce the total echo train length to minimize T₂ blurring. To minimize TE and maximize SNR for the center of k-space, centric reordering was implemented in the partition direction. EPI had FOV 98×32.8×12 mm³, with a matrix size of 132×44×8, for a nominal resolution of 0.75×0.75×1.5 mm³; TE was 24ms. Partial Fourier of 6/8 (with POCS) and GRAPPA acceleration of 2 were used in the in-plane phase encode direction. No acceleration or Partial Fourier was used in the partition direction. Note that while GRASE and EPI have similar FOVs and slice placement, read and phase encoding directions (Gr and Gp) were orthogonal for the two sequences to avoid aliasing along the phase encode direction in EPI. Although this leads to differing phase encode FOV and direction for the two sequences, both are optimal given the requirements and limitations for the two sequences. EPI requires a large phase encode FOV (with acceleration) to avoid wrap-around aliasing of the signal, while GRASE can use zooming to limit the phase encode FOV, minimizing T₂* contamination without the need for acceleration and the concomitant SNR penalty.

Figure 3:

Schematic of inner volume selection using 90° and α (≤180°) RF pulses, applied perpendicularly. This limits the FOV in the inplane (y) and partition (z) directions, allowing high-resolution acquisition with minimal echo train length, minimizing T₂* weighting in the image.

At 7T the blood nulling TI is around 1450ms (assuming blood T₁ = 2100ms (56)), which is very similar to blood arrival time in the motor cortex (while differences between arterial and venous blood T₁ have been reported, these small differences are unlikely to affect the VASO contrast so we assume a single blood T₁ (13)). To avoid the non-nulled blood arriving to the motor cortex during VASO image acquisition, an IR pulse with lower inversion efficiency was used for the GRASE acquisition. Specifically, a BIR4 pulse (57) was adapted to have a 71% efficiency (FA = 135°), which decreased the blood nulling TI to 1128 ms. We used TI = 1100 ms in these experiments, as slight deviations from blood nulling TI were shown to not substantially affect the VASO contrast (13). The following timing parameters were used: TI/TR = 1100/3000 ms, Acquisition Time (TA) = 12 min. The time between all imaging readouts was kept constant and was TR/2, as in previous VASO implementations to allow BOLD correction using interleaved blood-nulled and BOLD acquisitions (13,14). Fat Suppression was implemented for both sequences in the form of a gaussian saturation pulse (with spoiler) before each excitation pulse. The tSNR maps of all the sequences used can be seen in Supporting Information Figure S2. In general, tSNR for VASO was lower than BOLD for both sequences, and tSNR was higher in GRASE than in EPI (Supporting Information Figure S2A). Adding VFA to GRASE only reduced the tSNR by a small amount (Supporting Information Figure S2B).

Table 1 shows the imaging parameters for both sequences.

Table 1).

Imaging parameters for the two sequences used in the current study. Whilst nominal resolution was kept constant, other imaging parameters varied according to the sequence used.

	GRASE	EPI
FOV (mm)	99×25×12	98×32.8×12
Matrix	132×34×8	132×44×8
Phase Encode	AP	RL
In plane resolution (mm2)	0.75	0.75
Nominal slice thickness (mm)	1.5	1.5
Slices per slab	8	8
Partial Fourier	5/8 (Partition)	6/8 (Phase)
K-Space Partition Order	Centric	Linear
GRAPPA	-	2
TR (between IR pulses, ms)	3000	3000
TR (between image readouts, ms)	1500	1500
TE (ms)	48	24
TI (ms)	1100	1100
Excitation flip angle (degree)	90	26 – 90
Refocusing flip angle (degree)	165	-
BW (Hz/Px)	1052	1062
Inversion Pulse	BIR4	TR-FOCI

To increase SNR, the relatively predictable folding pattern of the “hand-knob” gyral pattern in M1 allowed thicker slices to be used when the slices were perpendicular to the central sulcus. Slice position was adjusted for each subject to be perpendicular to the thumb/forefinger region of their M1 hand-knob, based on a separate (0.8mm isotropic voxel size) MP2RAGE image. To mitigate B1+ inhomogeneities that can negatively affect the image quality of spin-echo sequences, and to help ensure a proper inversion pulse for blood nulling, a passive B₁⁺ shimming approach was adopted by placing high permittivity dielectric pads (containing a suspension of calcium titanate powder in water) around the head and neck (58,59). Transmitter reference voltage was adjusted for each subject by measuring the mean B₁+ values in the volume of interest with the dielectric pads applied, using a B1 mapping procedure (60).

Analysis

Pre-processing

Image volumes with VASO and BOLD contrast were separately corrected for motion using SPM12 (Functional Imaging Laboratory, University College London, UK), with the option of spatial weighting to optimize alignment over the motor cortex. A 4th order spline was used for resampling to minimize blurring (14).

At high field strengths, the positive BOLD signal outweighs the negative VASO signal caused by an increase in CBV (61), for both EPI and GRASE imaging. Dynamic division of the VASO and BOLD volumes was performed to remove residual BOLD contrast contamination from the VASO images, under the assumption that T₂* contrast is the same in images with both contrasts because they were acquired simultaneously in an interleaved fashion (11).

Functional and layer analysis

The functional activations were calculated as the difference between mean signal during the task and mean signal during rest, expressed as percent signal change per voxel, ignoring the initial 9s of each period to minimize the influence of transition periods. This method has been shown to provide results that are easier to interpret than methods using inferential statistics, which can be affected by laminar differences in noise and hemodynamic response function (HRF) shape (14). Additional GLM analysis was also done using the FSL FEAT toolbox (62).

Functional data were used to create the WM and CSF borders for the laminar analysis to avoid issues of distortion, registration and interpolation (63). Functional VASO images can be used to generate a T₁-weighted anatomical image (based on the inverse signal variability in the timeseries) that provides good contrast between GM and WM (Figure 4A), which were used as anatomical references for grey matter boundaries in the EPI images (14,64). For GRASE scans the mean BOLD images showed greater tissue contrast, so these images were used for boundary identification. To create masks of cortical depths, images were 5 times upsampled (cubic interpolation) and the GM boundaries with WM and CSF were manually delineated on the anterior bank of the central sulcus, including the hand-knob region of M1 (Figure 4B & C).

Figure 4:

Demonstration of the steps for forming surfaces at different cortical depths directly on the functional images.

A) T₁-weighted EPI from VASO data.

B) GM/WM (orange) and GM/CSF (red) boundaries delineated on the data.

C) Cortical GM mask created.

D) Equidistant cortical depths calculated within GM mask.

E) ROIs restricted to lateral bank of M1 hand-knob.

Cortex was divided into 21 equidistant depths in the software suite LAYNII and the functional analyses were performed within each, as per previous work (13,14). Figure 4D shows an example of 21 cortical depths overlaid over the T₁-weighted image obtained from the VASO data. These 21 depths could be used to plot signal change across cortex, as well as to do selective smoothing restricted to cortical depth, which aided in visualizing the expected bimodal pattern of activation across depth (14,65). Activation at different depths was measured in an ROI confined to the lateral side of the hand-knob in M1 (Fig 4E). This is the evolutionary older portion of the hand-knob, located less deep in the central sulcus (aka. ‘old’ M1, rostral M1, or BA4a) (66), which lacks the cortico-motoneuronal cells and is the location of the double peak feature investigated in previous studies (14). Sensitivity and specificity were examined by calculating the average signal change across layers (sensitivity), and by the slope of a line fit to each depth profile that indicated the level of pial bias (an inverse measure of specificity) (14,67), in line with several models describing a (vein driven) linear signal trend across cortex for gradient-echo sequences (68,69). A fit including a bimodal function (sin²) was also tested. The effect of sequence type on sensitivity and pial bias was tested for significance with a repeated measures ANOVA followed by post-hoc analyses.

Point Spread Function (PSF) Simulations

To investigate the signal behaviors of GM tissue in the EPI and GRASE sequences, Bloch equation simulations were performed by numerical application of 3×3 rotation and relaxation matrices, followed by averaging the signal intensity over spin isochromats at each echo time under different regimes of constant flip angle (CFA) and VFA. The simulation parameters were identical to imaging parameters summarized in Table 1, including the inversion pulse. The CFA used were 16° for 3D EPI and 165° and 180° for 3D GRASE, respectively. The VFA used for EPI were the same as in the current experiment, and for 3D GRASE were 64°-110° (calculated by solving an inverse solution of the Bloch equation) (48). A T₁ of 1900ms and a T₂ of 60ms were assumed for GM. PSF for all simulations was numerically estimated by mapping the simulated GM signal evolution into partition direction according to the reordering (centric for GRASE, linear for EPI) to create an MTF, and using the magnitude of inverse Fourier transformation of the MTF as an estimate of the imaging PSF. To help better illustrate the PSF at a sub-voxel level, the MTF was also zero padded prior to the inverse Fourier transform, corresponding to an upsampling (via sinc interpolation) of the PSF.

Results

Task-related signal change

Figure 5 shows the mean relative signal change over 12 block repetitions for 6 volunteers (Error bars = standard error of the mean (SEM). EPI- and GRASE-BOLD signal (green and red lines, respectively) increases with activations; EPI- and GRASE-VASO signal (blue and black lines, respectively) decreases, as an increase in CBV corresponds to a decrease in VASO signal. EPI-BOLD shows the largest relative signal change (6.5% ±1.3), GRASE-BOLD and EPI-VASO show similar amplitudes (with opposite signs) (3%±0.48, 3%+/−0.43), and GRASE-VASO shows the lowest amplitude (2%±0.4). Additional data from a separate session where 4 of the subjects were rescanned, and data from an additional experiment where subjects were scanned using VFA GRASE, are shown in Supporting Information Figure S3.

Figure 5:

Average timecourse in hand-knob region of M1 across stimulus cycles for EPI-BOLD (green), EPI-VASO (blue), GRASE-BOLD (red) and GRASE-VASO (black). BOLD signal change is positive during finger tapping, VASO signal change is negative relating to an increase in CBV. Errorbars = SEM across subjects (N=6).

Functional signal change across cortical depths

Figure 6 shows example signal change maps for EPI and GRASE in a single subject, for both BOLD and VASO contrasts. Unsmoothed and surface-smoothed maps are shown for each contrast/readout comparison. Note that owing to differences in contrast mechanisms between BOLD and VASO, signal change maps are differently thresholded. Additional signal maps for additional subjects are shown in Supporting Information Figure S4. In addition, z-maps from a GLM run on the data using FSL FEAT (62) are included in Supporting Information Figure S5.

Figure 6:

Signal change maps, masked by GM, for an example subject for EPI-BOLD and -VASO, and GRASE-BOLD and -VASO. EPI-BOLD maps show highest values around the pial surface, whereas EPI-VASO and GRASE-BOLD both show two peaks at deep and superficial depths. Note that BOLD and VASO are thresholded differently due to the different contrast mechanisms involved. Black contour indicates the region from which depth profiles were calculated.

Figure 7 shows the average depth profiles for each scan type across subjects. Sensitivity is defined as the mean percent signal change across depth for each scan type, and specificity defined as the (inverse of) pial bias, as measured by a slope fit to each profile across depth (14). Individual depth plots are shown in Supporting Information Figure S6. Each subject shows the expected surface bias for EPI-BOLD, and evidence of double peaks for EPI-VASO. Double peaks of varying clarity are also seen in the GRASE-BOLD and -VASO data. To ensure that the depth profiles did not arise from the choice of interpolation, depth profiles were redrawn on data upsampled using nearest neighbor (NN), linear and cubic interpolation, showing only subtle differences at the group level (Supporting Information Figure S7). To examine whether using a different measure of activation that included the effects of imaging and physiologic noise instead of percent signal change affected the profiles, profiles were replotted using z-scores (Figure 7, right). Results from the second experiment where subjects were scanned using GRASE-VASO sequences with and without VFA indicated little overall difference in terms of specificity (Supporting Information Figure S8).

Figure 7:

Average depth profiles for signal change between superficial (toward CSF) and deep (towards WM) depths for Percent Signal Change (left) and Z-Score (right). Errorbars = SEM across subjects (N=6). Dotted lines indicate fitted lines used for evaluating pial bias for each sequence.

Figure 8 shows comparisons of sensitivity and pial bias (inverse specificity) for each contrast. The effect of sequence type on sensitivity and pial bias was tested for significance with a repeated measures ANOVA with post-hoc analyses. There was a significant effect of sequence type on sensitivity (F_3,15 = 20.4, p<0.0001) and pial bias (F_3,15 = 9.29, p<0.01). Post hoc analyses (Bonferroni corrected) revealed a significant difference in sensitivity between EPI-BOLD and GRASE-BOLD (p<0.01) and GRASE-VASO (p<.01), a significant difference in sensitivity between EPI-VASO and GRASE-VASO (p<0.05), and a significant difference in pial bias between EPI-BOLD and GRASE-BOLD (p<0.05). Repeating the fit with a bimodal component (sin² function) showed that while the fit for all sequences was improved by including the bimodal component, EPI-BOLD was still dominated by the linear component showing the large impact of the pial bias of this sequence (Supporting Information Figure S9).

Figure 8:

Comparison of sensitivity and pial bias amongst the four sequences for Percent Signal Change. Post-hoc analysis indicates that EPI-BOLD (green) shows significantly greater sensitivity than GRASE-BOLD (red) and GRASE-VASO (black) and EPI-VASO (blue) shows greater sensitivity than GRASE-VASO. In addition, EPI-BOLD shows significantly greater pial bias than GRASE-BOLD.

Simulations of the signal variation across the echo trains for EPI and GRASE (Figure 9A), and the resulting variation in signal across k-space partitions (Figure 9B) yield estimates of the PSF in the slice direction for the two acquisitions (Figure 9C). Due to the signal variation caused by T₂ decay in GRASE, the PSF in the slice direction for GRASE is around 2.6–2.7 times that for EPI (GRASE 180 FWHM = 3.22, EPI FWHM = 1.19), as the latter sequence maintains a flat signal profile across the echo train due to the VFA approach used in this acquisition (EPI CFA FWHM = 1.2). Concomitant with previous studies (48), using a constant but less than 180° refocusing pulse yields a very slight improvement in blurring (GRASE 165 FWHM = 3.14), albeit much less than that from using VFA (GRASE VFA FWHM = 2.12). Plotting the relevant PSFs in terms real and imaginary components (Figure 9E-G) shows the same pattern of larger PSFs for GRASE.

Figure 9:

Point Spread Functions (PSF) in the slice/partition direction for EPI and GRASE under different regimes of constant flip angle (CFA) and variable flip angle (VFA) A) Simulated signal across the echo train for GRASE with a CFA (180°, blue; 165°, orange) and VFA refocusing (yellow) train to compensate for or reduce T₂ decay, and EPI without (purple) and with (green) a VFA excitation component to compensate for T₁ recovery. The use of VFA results in a flatter signal profile across the echo train. B) The resulting Modulation Transfer Function (MTF) for the GRASE sequences after centric reordering, and the 3D EPI sequences after linear phase encoding. C) The resulting Magnitude Point Spread Functions (PSF) from the respective (zero padded) MTF, corresponding to upsampling of the PSF. EPI (with and without VFA) has a narrower PSF than GRASE with a constant flip angle. The use of VFA refocusing in GRASE narrows the PSF and would mitigate blurring on the slice axis. D-F) The real and imaginary components of the PSF for CFA GRASE, VFA GRASE and VFA EPI, showing the origin of the sidelobes in the PSF.

Discussion

Previous studies have demonstrated the utility of CBV-based VASO imaging for high-resolution, laminar fMRI, being less weighted towards surface vasculature than GE-BOLD with the latter’s T₂*-based contrast mechanism. A similar suppression of signal from surface vasculature has been shown for GRASE, raising the question of how pial bias compares between VASO and GRASE. We extend previous work combining GRASE and VASO at standard resolutions to UHF, sub-millimeter imaging for laminar analysis.

Comparing relative signal changes in M1 for each contrast/readout (Figure 5), standard EPI-BOLD showed the highest levels of signal change, on the order of 6%, with GRASE-BOLD showing lower levels of around 3%, in line with previous measurements (10,70). The levels of signal change for EPI-VASO were ~3%, similar to previously seen changes using EPI (13,14). Signal change in GRASE-VASO was lowest at around 2%.

The lower signal change for GRASE-BOLD compared to EPI-BOLD is expected given that EPI BOLD contains susceptibility contrast changes from both large and small vessels, whereas GRASE primarily has signal change from only small vessels. In addition, GM signal (and hence VASO contrast) is maximized with the VASO acquisition used here when the overall TR is kept minimal. In this case the time between image acquisition was 1.5s, which was insufficient for full recovery of the GRASE signal in GM using full 90 RF excitation and may have led to suboptimal BOLD imaging using GRASE. Without the need for a VASO acquisition component, longer TRs of 2 to 3s seconds would increase the delay between readouts, and allow greater signal recovery for GRASE BOLD imaging (with the same overall number of samples for BOLD imaging as in the VASO acquisition). These kinds of acquisitions have been utilized at 7T in mesoscale fMRI studies (8–10,35). This lack of overall image SNR may have also led to the lower VASO contrast seen using GRASE-VASO compared to EPI-VASO, despite equivalent levels of signal change observed in previous implementations (42). In theory a single TI, defined by the single excitation pulse for each single-shot GRASE readout, would have led to more consistent blood nulling across k-space (42,43). This could potentially yield a more accurate measurement of CBV (after BOLD correction using the second image readout) in the GRASE-VASO signal compared to EPI-VASO, in which the multiple excitations for k-space partitioning lead to inversion times that are not at the blood nulling time for some parts of k-space. However, the lower signal in GRASE-VASO obscured these smaller effects of imperfect blood nulling compared to segmented EPI-VASO. The timing parameters (TI/TR) for GRASE-VASO were based on those previously optimized for EPI-VASO (14), and it may be the case that a different implementation could lead to similar amplitudes for EPI- and GRASE-VASO, as seen previously at 3T. The short TR and subsequent necessity of a reduced FA for refocusing pulses also reduced the overall SNR for both GRASE-BOLD and GRASE-VASO imaging, and optimized parameters for GRASE-BOLD imaging sequence alone, without being incorporated into VASO acquisition, would further improve sensitivity and the amount of pial bias.

Another reason for the difference in relative signal change between EPI-VASO and GRASE-VASO might be from a different signal evolution along the finite readout duration. The EPI-VASO readout consists of multiple excitations along the inversion-recovery relaxation. While the k-space center is acquired at the TI of the blood nulling time, outer k-space segments (collected before and after the blood nulling time due to the linear k-space ordering) can contain residual blood signal. Dependent on the water permeability between intravascular and extravascular space, residual blood signal can accumulate in the extravascular space during the readout and, thus, amplify the VASO contrast. This extra-vascular perfusion component is to some extent balanced out by exchange in the other direction, but simulations show that the net result is an amplification of CBV estimates (11,71). This perfusion-dependent signal weighting is not to be confused with potential CBF-dependent inflow effects of fresh, not-inverted blood (described in (72)). Such inflow would reduce the VASO contrast, while the permeability effect here refers to perfusion related CBV amplification. While this effect across the short readout train used for EPI-VASO should be minimal, the fact that GRASE-VASO uses only one excitation pulse means that the nulled blood signal at excitation represents a true snapshot along the inversion-recovery relaxation, without permeability-related amplifications. Thus, in theory, GRASE-VASO should constitute a more accurate/quantitative CBV signal. Non-zero longitudinal magnetization in other tissues however will potentially lead to a complex signal evolution during the following train of refocusing pulses.

It is also worth noting that based on the different spin-history in EPI-VASO and GRASE-VASO, they are differently sensitive to potential contamination of dynamic CSF volume changes (73,74). We do not believe that such hypothetical CSF contaminations can explain the different signal changes in our results for the following reasons. First, we previously found that for local activation in the motor cortex during finger tapping, the CSF volume change is negligibly small compared to global systemic respiration tasks (13). Second, such a CSF contamination should be isolated to partial voluming in superficial layers (25). The profiles in Figure 7, however, show that the GRASE-VASO has a reduced signal change across all cortical depths. Thirdly, the longer T₂ of CSF leads to a narrower PSF arising from this tissue (7), further minimizing any potential contaminations

Maps of signal changes for the four sequence types in an example subject (Figure 6) and the average profiles of signal change across depth for the four contrasts (Figure 7) mirror the pattern seen in the time course (Figure 5) results. EPI-BOLD has overall higher signal change, but the larger signal change is biased towards superficial cortical depths, as seen previously in humans (3,10,13,14). This indicates that EPI-BOLD is mostly influenced by the signal change in large draining veins at the pial surface. EPI-VASO shows little to no surface bias and shows peaks at both superficial and deep cortical depths, as seen in previous studies (13,14). Similarly, GRASE-BOLD does not show a superficial bias, in line with previous studies (7,10), and also shows the double peak pattern (Figure 7) visible in EPI-VASO. GRASE-VASO shows the same pattern to some extent although with higher levels of noise, mirroring the lower CNR seen in the time course plots (Figure 5).

The double peaks observed in the GRASE-BOLD data are of lower amplitude than those observed with EPI-VASO (Figure 7). The depth profiles in both GRASE-BOLD and -VASO for individual subjects (Supporting Information Figure S4) demonstrate overall larger variations in amplitude compared with those using EPI-VASO. In addition to the suboptimal SNR arising from the timing parameters used, this variability may be due to the broader point spread function (PSF) in the slice direction for GRASE, and dependent larger partial volume effects in cases where the slice placement is suboptimal, i.e. not completely perpendicular to the cortical hand knob region of M1. Repeating the experiment with a GRASE-VASO acquisition with and without VFA seems to indicate the PSF is not the driving force behind this difference, as VFA GRASE did not significantly improve the profiles obtained (Supporting Information Figure 8A). A second issue is that cortical surfaces were defined separately for the EPI and GRASE scans in each subject to analyze data in native functional space, and in some cases the GM/WM contrast was less clearly defined for GRASE data. This could have led to inconsistencies in placing the GM/WM boundary when defining surfaces, making the peaks in the group average profile less distinct even when seen in individual depth profiles. We also chose to use percent signal change to assess sensitivity and pial bias, based on previous work that used this metric (7,10,14), and to avoid the effect of potential HRF differences across layers (14). However, this neglects the issue of imaging and physiological noise, which can contribute to the sensitivity of a given method (75). Repeating the analysis based on the z-score from a GLM rather than percent signal change (Figure 7, right) showed similar profiles for EPI-BOLD and GRASE-BOLD, with z-scores for the latter approaching those of the former despite lower signal change, likely due to lower variability in the GRASE sequence (7). The profiles for the two VASO sequences both showed low z-scores, with EPI-VASO in particular no longer showing greater sensitivity than GRASE-VASO, and showing an altered profile that shows the double peaks less clearly. Replotting the second experiment using z-score showed that VFA GRASE-BOLD had lower z-scores than GRASE-BOLD without VFA (Supporting Information Figure 8) while still demonstrating the expected double peak profile, as expected from previous work (7).

We compared sensitivity (mean signal change across depths) and specificity (pial bias, measured by the slope of a linear fit to the signal change across depth) (14) for each contrast type (Figure 8), and found a significant effect of sequence type on both sensitivity and pial bias, with post-hoc analyses revealing EPI-BOLD to have a higher sensitivity than other sequences, but also a higher pial bias (i.e. less specificity). It should be noted that the definition of sensitivity and specificity (pial bias) used here are not entirely separable. For example, the case of absolutely no signal change would be treated as highly specific (i.e. a flat line), so for this analysis sensitivity and specificity should only be considered together. This definition of specificity (i.e. degree of pial bias) has been used previously in studies examining laminar responses using fMRI (10,14,67), and a previous comparison with alternative measures of specificity found unchanged results (67). An additional analysis adding a bimodal function (sin²) to the fit showed that while this improved the fit in all cases, the EPI-BOLD fit was still dominated by the linear component, demonstrating the large effect of the surface bias in this case (Supporting Information S9). Many different methods of assessing sensitivity and specificity exist, including taking into account the orientation (76) and spatial arrangement (77) of vessels at different depths, and making use of the temporal information available in the signal from different depths (78), or testing the functional specificity of the responses to set of stimuli or stimulus properties (10,70). Each method has some drawbacks and assumptions about the signal being examined, and although the slope of the linear fit (i.e. degree of pial bias) may be a simplified way of examining specificity, it does in this case illustrate the large pial bias in EPI-BOLD that is not present in the other sequences, as seen in previous comparisons (10,14,70) and in agreement with models of the effect vasculature (68).

While it should be noted that key imaging parameters (acceleration, zooming, phase encode direction etc.) differ across the two imaging readouts used (EPI and GRASE), each set of parameters were optimized for the sequence in question to obtain high in-plane resolution for laminar imaging. The EPI sequence parameters were optimized based on a previous experiment in M1 (14), and these parameters would not be optimal for a GRASE sequence, so GRASE was optimized separately (within the constraints of a VASO sequence). Thus, the experiment was not to compare two sequences identical except for their contrast mechanisms, but to compare two sequences independently optimized for mesoscale imaging.

Previous comparisons of VASO and SE-BOLD showed the latter still having a residual surface bias when compared to VASO, albeit less than that seen in GE-BOLD (14). This seems contradictory to the lack of surface bias seen using a SE sequence (GRASE) in the current study, however the difference likely arises from the use of much shorter EPI echo trains in the zoomed GRASE sequence than SE-EPI in the previous comparison. SE-EPI is known to have large vessel contributions arising from T₂* contamination in the extended echo train at high resolutions (16,21). The use of inner volume zoomed GRASE has been shown to decrease surface bias when compared with non-zoomed SE-EPI (7), and the decreased pial bias seen with GRASE compared with that previously demonstrated with SE-EPI replicates this observation. Another difference between SE-EPI and GRASE is the T₁ contrast contribution from STE in GRASE, not present in single refocusing SE-EPI, which increases with reduced flip angle. In our experiments a nominal flip angle of 165° instead of 180° refocusing pulse was necessitated to stay within SAR limits using the 1.5s effective TR between GRASE readouts, but simulations of the signal evolution under a 180° and 165° refocusing regime indicate that there is little difference between the two in terms of STE contribution (Supporting Information Figure S1B).

VASO is also not completely independent of macrovascular biases towards the pial surface. As previously discussed in (Figure 8 of (13),(79)), the larger vascular density of diving arterioles and micro-vasculature in superficial and middle layers can result in higher signal changes compared to deeper layers.

The results of this work show that spatial specificity of both GRASE-VASO and GRASE-BOLD is sufficient for laminar specific imaging, yielding two distinct functional activation peaks across the depth cortex in M1 (Figure 7), in agreement with underlying input and output-driven activity associated with a finger-tapping task.

The use of different image contrasts for laminar fMRI in humans has been of increasing interest since mesoscale resolution became achievable. The weighting towards pial vessels shown for GE-EPI makes the straightforward resolving of BOLD signals from specific laminae difficult (3,10), necessitating that superficial depths, non-specific voxels, or voxels containing veins be excluded during analysis (2), or necessitating unique experimental designs (80,81) ,analyses (8,70) or modelling efforts (68,82,83) to begin to resolve layer specific signals. The use of SE contrast at high field strengths to suppress signals from large veins has been combined with 3D imaging in GRASE (5) to demonstrate mesoscale fMRI without this bias towards the pial surface (7,10). GRASE has been used to demonstrate consistent selectivity for a given stimulus property across depth (the hallmark of cortical columns) in various cortical areas (9,36,70), and to demonstrate certain cortical computations restricted to certain cortical depths (8,35). The current study is the first to show two distinct activity peaks in human M1 using BOLD fMRI, which has been used as a hallmark of laminar specificity when assessing different contrasts for laminar fMRI (14).

Previous work suggests that VASO (12) acquisition methods for fMRI have an optimal tradeoff of functional sensitivity and spatial specificity (11,14). However, VASO imaging requires implementing an adiabatic IR RF pulse, which has to provide inversion with decreased efficiency [e.g. a BIR4 (57) or TR-FOCI (84) pulse] and increases the acquisition time by blood nulling time, on the order of 1 s. It also requires a very careful timing implementation, so that no fresh (non-inverted) blood flows into the imaging volume, which requires the arrival time and T₁ of blood in a given brain area to be known. While variable blood T₁ across subjects has been reported, the effects of these small variations in T₁ are unlikely to have a large effect on the VASO signal (13). In addition, there is residual BOLD contrast in the blood-nulled images, especially at ultra-high-fields, which is of the opposite sign and frequently of higher amplitude than the negative VASO signal. This problem can be alleviated using the BOLD-corrected VASO (11) method, where two images with and without blood-nulling are required, and the latter, purely BOLD weighted image is used to correct the residual BOLD signal in the blood-nulled image. However, this requires additional image post-processing and creates the problem of reduced temporal resolution of the VASO acquisition as compared to the BOLD images alone. Further development and sequence parameter optimization of GRASE alone could have advantages to VASO as it would eliminate the use of complex sequence timing with inversion pulses and the inefficiency of acquiring a double readout for BOLD T₂* correction. A recent development of compressed sensing (CS) GRASE to increase slice coverage is a promising further development of the methods used here (85,86).

Although the sensitivity and lower pial bias of GRASE demonstrated here make it a good candidate for mesoscopic imaging, it is not without drawbacks. As noted above, the PSF in the slice direction is increased for GRASE (Figure 9) due to the T₂ decay across the echo train which can be mitigated to an extent using VFA on the refocusing pulses (48), at the expense of some overall SNR. These differences in PSF between the used sequences could affect the sensitivity and specificity as measured in this paper, for example through partial volume effects. However, in this particular experiment, the larger PSF in the slice direction for GRASE will have less of an effect than it might otherwise owing to the nature of the neural anatomy under investigation, where thicker slices were placed orthogonal to the central sulcus and laminar activity was resolved inplane. In more convoluted areas of cortex where resolutions closer to isotropic are required, methods to mitigate through plane blurring in GRASE such as VFA (48) or CS (85,86) may prove critical, although the curvature of cortex within an imaging volume may mitigate the effect of through-plane blurring on cortical depth profiles (7). The robust functional results obtained using GRASE, despite its larger PSF, indicate that larger through-plane voxels can be used in cases where the slice plane can be placed orthogonal to the area of interest. For example, the EPI data could be acquired with an even larger nominal slice thickness and/or a shorter acquisition time without a loss in sensitivity or specificity.

The question still arises as to whether these T₁ contributions from STE in GRASE increase the risk of inflow artifacts (87,88). Although on the one hand the longer T1 of blood at high fields (56) may lead to an increased chance of inflow artifacts, the simultaneous shortening of blood T₂/T₂* is likely to counteract this. In addition, as GRASE is a 3D sequence, with orthogonal slab excitation and refocusing volumes, this should also reduce the potential impact of inflow artifacts. Additionally, only short (5 refocusing pulses) SE trains were used in these GRASE experiments, giving less opportunity for STE magnetization to build up in the later echo train.

In addition to the inflow effects commonly seen in BOLD images, all VASO sequences are at risk of artifacts due to the inflowing of uninverted blood to the imaging volume, e.g. in areas of very short arrival time such as large arteries (52), potentially reducing any observed VASO signal change. The use of shorter TIs with adjusted inversion efficiency, as performed here, can help avoid these artifacts.

Conclusion

These results demonstrate the feasibility of using a 3D-GRASE readout at ultra-high field for laminar fMRI. The sensitivity and specificity of laminar fMRI achieved with BOLD based 3D-GRASE is similar to CBV-based VASO with a 3D GE-EPI readout. GRASE-BOLD is a CPMG spin echo train sequence in which inner volume zooming further reduces T2* contrast by shortening the EPI echo trains, and taken together, largely removes the signal contribution of pial draining veins, providing laminar specificity with BOLD contrast. The combination of the VASO technique with a 3D-GRASE readout in effect combined two independent approaches to suppress pial vein contributions, however this did not improve sensitivity or specificity, indicating that either standard GRASE with T2-weighted BOLD contrast, or VASO using an EPI readout are more optimal for laminar fMRI applications.

The ability to specifically measure neuronal signal in cortical layers provides a unique approach to studying circuitry in the human brain where different cortical layers act as inputs from, or send outputs to other cortical areas. By mapping this mesoscale organization in brain circuitry, cortical laminar fMRI could be used to study the brain’s directional hierarchy in long-range circuits between the hundreds of distinct brain areas and the microcircuits within cortical layers. Therefore, laminar fMRI may also play an essential role in understanding the dysfunction of brain circuits affecting higher-order brain functions.

Supplementary Material

supp info

A) Relative SNR and SAR for GRASE-VASO with increasing refocusing flip angles from 90° to 180°. B) Signal evolutions for 180 and 165 GRASE sequences, showing both the signal with T2 decay and the Stimulated Echo for 165°.

A) tSNR maps from 6 subjects for the 4 different contrasts in Experiment 1. B) tSNR from 6 subjects in Experiment 2. Similar to earlier VASO measurements (47) the largest source of variance in measurements is the inter-participant variance. This may be due to the participant specific slice alignment (tilting and proximity to RF coils), which has associated variability in gradient delays (ghosting) and GRAPPA g-factor penalties. One difference across studies might be that the vendor supplied POCS partial Fourier reconstruction algorithm used in a 2018 paper (47) was faulty and later fixed (private communication L.H.). The current VASO data using the correct POCS reconstruction chain, therefore may make images sharper but also noisier than the earlier images.

A) Mean percent signal change (left) for the four sequences in 6 subjects, also shown scaled to each curve’s maximum. B) Same data from a follow up session in 4 subjects. C) Data from a second experiment using GRASE and VFA GRASE.

Smoothed and raw signal change maps for individual subjects for EPI BOLD/VASO and GRASE BOLD/VASO. Note that BOLD and VASO are thresholded differently due to the different contrast mechanisms involved.

Smoothed and raw Z-Score maps for individual subjects for EPI BOLD/VASO and GRASE BOLD/VASO.

Depth profiles for individual subjects for EPI (left column) and GRASE (right column) for BOLD and VASO.

Depth profiles using 3 different upsampling methods.

Depth profiles using GRASE VASO with and without VFA (N=6) for % Signal Change (left) and Z-Score (Right)

Fitting the depth profile for % Signal Change and Z-Score when including a bimodal component (sin²). Top row shows depth profiles for % signal change (left) and z-score (right) with a fit including a bimodal component (dashed lines). Bottom row shows variance explained by the linear and bimodal component of the fit.