Sensitivity limitations of high-resolution perfusion-based human fMRI at 7 Tesla

Abstract

The study of the brain’s functional organization at laminar and columnar level of the cortex with blood oxygenation level dependent (BOLD) functional MRI (fMRI) is affected by the contribution of large veins downstream from the microvascular response to brain activity. Blood volume- and especially perfusion-based techniques may reduce this problem because of their reduced sensitivity to venous effects, but may not allow the same spatial resolution because of smaller signal changes associated with brain activity. Here we investigated the practical resolution limits of perfusion-weighted fMRI in human visual stimulation experiments. For this purpose, we used a highly sensitive, single-shot perfusion labeling (SSPL) technique at 7 T and compared sensitivity to detect visual activation at low (2 mm, n=10) and high (1 mm, n=8) nominal isotropic spatial, and 3 s temporal, resolution with BOLD in 5½-minute-long experiments. Despite the smaller absolute signal change with activation, 2 mm resolution SSPL yielded comparable sensitivity to BOLD. This was attributed to a superior suppression of physiological noise with SSPL. However, at 1 mm nominal resolution, SSPL sensitivity fell on average at least 42% below that of BOLD, and detection of visual activation was compromised. This is explained by the fact that at high resolution, with both techniques, typically thermal noise rather than physiological noise dominates sensitivity. The observed sensitivity loss implies that to perform 1-mm resolution, perfusion weighted fMRI with a robustness similar to BOLD, scan times that are almost 3 times longer than the comparable BOLD experiment are required. This is in line with or slightly better than previous comparisons between perfusion-weighted fMRI and BOLD. The lower sensitivity has to be weighed against the spatial fidelity advantages of high-resolution perfusion-weighted fMRI.

1. INTRODUCTION

To date, functional MRI (fMRI) studies have relied heavily on blood oxygenation-level dependent (BOLD) contrast [1], which has proved exquisitely sensitive to the blood flow changes elicited by the neuro-vascular response to brain activity changes. This sensitivity has been exploited to obtain functional images with spatial resolutions of 1 mm and below, with the goal of resolving activity patterns at the levels of cortical layers and columns [2–5]. Nevertheless, remaining limits in the way of this goal are the resolution loss and mis-localization associated with BOLD effects in veins downstream from the activated area [6]. This venous contribution limits the spatial accuracy of typical fMRI experiments to a few millimeters [7], insufficient for reliably pinpointing the source of activation on the scale of cortical layers or columns. Intra-cortical and pial veins furthermore bias the response profile as a function of cortical depth in BOLD fMRI experiments [8], another drawback of this approach, albeit this can presumably be corrected for in post-processing [9].

While special stimulation tasks [10] and processing approaches [9, 11] may ameliorate this problem somewhat, alternative fMRI contrasts less sensitive to venous drainage may be desirable for high resolution studies [12]. For example, perfusion [13], blood volume (BV) [14] and spin-echo (SE) BOLD [15] based fMRI methods have been demonstrated to improve spatial accuracy [2, 6, 12, 14]. However, these methods suffer from reduced sensitivity when compared to gradient-echo (GE) BOLD, e.g. [16], primarily because of the smaller absolute signal changes associated with these methods. While this may be overcome by extensive signal averaging, this typically leads to impractically long scan times.

One way to improve the sensitivity of perfusion-based fMRI is the use of a reference-less acquisition, as proposed with the single-shot perfusion labeling (SSPL) technique [17]. SSPL can improve the sensitivity to detect perfusion changes by two-fold over conventional techniques [18, 19], while sacrificing the ability to extract quantitative values. However, since fMRI typically attempts to detect the effects of blood flow (and/or volume) changes rather the flow parameters themselves, this may not be detrimental.

SSPL relies on a dual-inversion-based suppression of background signal [20, 21], which reduces confounding contributions from BOLD contrast and physiological noise, both of which are proportional to the background signal level. Typically, 5–20-fold reductions in background signal are achieved, rendering the remaining signal variability dominated by perfusion changes [17, 22]. Thus, while activation may generate relatively small signal changes in SSPL (as compared to BOLD), its favorably reduced levels of physiological noise may result in an overall adequate sensitivity for high resolution fMRI. In addition, the sensitivity of SSPL may also be competitive with cerebral blood volume (CBV) based techniques, as according to Grubb’s law, fractional perfusion changes substantially exceed fractional BV changes [23].

The aim of this work was to investigate the feasibility of SSPL for 1 × 1 × 1 mm³ (1 μl) resolution functional imaging, when implemented with state-of-the-art hardware. We anticipated that the sensitivity penalty of employing a notably higher resolution, such as 0.7³ mm³ or even 0.5³ mm³, would be too large for the relatively brief scans used here. (Notably longer scans come at an additional cost, such as increased motion sensitivity.) For this study, visual fMRI experiments were performed in humans at 7 Tesla, comparing fMRI performance against BOLD at two isotropic spatial resolutions of respectively 2 mm and 1 mm.

2. MATERIAL AND METHODS

2.1. The SSPL pulse sequence

Our SSPL fMRI technique is based on a multi-slice gradient-echo echo-planar imaging (EPI) image acquisition preceded by a perfusion preparation based on two slice-selective inversion pre-pulses [17]. By properly adjusting the timing (inversion times TI1 and TI2) and selection thicknesses of these two inversion pulses (IR1 and IR2), stationary magnetization can be suppressed, while magnetization of inflowing spins can be preserved (Supplementary Material, Figure S1). To minimize BOLD signal, the sequential echo planar imaging (EPI) slice acquisitions are performed with minimal echo time (TE). Remaining BOLD effects may either add to or subtract from the perfusion signal, depending on their relative signs. This can be controlled by properly adjusting the time range over which the slices are acquired (Fig. 1b in [17]). A pulse sequence diagram and the resulting evolution of magnetization (using T₁ values at 7 T taken from [24]) are shown in the Supplemental Material, Figure S1.

2.2. MRI experiments

Sequence optimization and fMRI experiments were performed on a Siemens Magnetom 7T 830 mm diameter horizontal-bore whole-body scanner (Siemens, Erlangen, Germany) with SC72CD gradient set and 32-channel Nova Medical receive coil (Nova Medical, Wilmington, MA, USA). We acquired data from 9 human healthy volunteers (5 m, 4 f, aged 22.7–58.5 years, some scanned more than once), which provided informed consent under a protocol approved by our institutional review board. An in-house developed pulse sequence was used that allowed conventional BOLD-EPI as well as SSPL-EPI and a dual-inversion flow-sensitive alternating inversion-recovery (FAIR)-like EPI (see below for details).

To optimize the timing (TI1 and TI2, Fig.1 in [17] and Fig S1) of the SSPL inversion pre-pulses for 7 T, data were acquired using 120 × 90 acquisition matrix size and 240 × 180 mm² field-of-view (FOV), yielding nominal 2³ mm³ (8 μl) resolution voxels (subsequently referred to as ‘low-resolution’). These experiments (n=10) used rate-2 sensitivity encoding (SENSE) [25], 3 s repetition time (TR) and 17.5 ms TE with a 250 kHz acquisition bandwidth. Echo spacing was 0.67 ms, for an overall readout duration of 30.52 ms. Total scan time was 330 s for the 110 acquired repetitions. Slice spacing was 25% of the 2 mm thickness, or 0.5 mm. The slice stack was placed parallel to and centered around the calcarine fissure.

The high-resolution scans (n=8) were acquired with 240 × 180 matrix size at 1³ mm³ (1 μl) resolution. Scan parameters kept similar to the low-resolution experiments above. Notable differences were a SENSE rate of 3, 38 ms TE, 1.16 ms echo spacing and 70.44 ms read-out duration. Slice spacing was increased to 150% of the 1 mm thickness, or 1.5 mm, so that low-resolution and high-resolution slices were centered in the same locations for scan sessions in which both where acquired. (This was done to potentially allow direct comparison of low-resolution and high-resolution data. There was no technical limitation requiring increased slice spacing.) In two high-resolution sessions, 210 repetitions were acquired, for a total scan time of 630 s. In all other experiments 110 repetitions were acquired, like to the low-resolution experiments.

In the original SSPL implementation, only a single slice was acquired. Here, a multi-slice implementation was investigated in which several imaging slices were acquired around the nominal second inversion time TI2. In the data shown, 5-slice SSPL was acquired. In the same 3-second TR, BOLD-EPI with similar resolution and echo time allowed the acquisition of 45 slices for low-resolution, and 33 slices for the high-resolution experiments. Two slice-selective hyperbolic secant inversion-recovery (IR) pulses (10240 μs duration; 800 Hz amplitude; 30- and 120-mm thickness for IR1 and IR2, respectively) were used for the inversion pulses in SSPL.

To quantify perfusion changes measured with SSPL and compare them to baseline perfusion, we also performed experiments (n=4) with a modified version of SSPL that on alternate volumes replaced the narrow-slab selective inversion with a non-selective one. This yielded a dual-inversion FAIR-like pulse sequence (along the lines of [20, 21]). Subtracting subsequent volumes in the resulting data allows perfusion quantification, as in conventional FAIR. Parameters used were otherwise identical to low-resolution SSPL.

During a preparation phase at the beginning of each scan, a noise measurement was performed, and an internal reference to aid in SENSE reconstruction was acquired. In SSPL, the reference volume was acquired without inversion pulses, giving it the same contrast as the corresponding BOLD data. From these pre-scan data, one coil-combined magnitude volume without IR was reconstructed, referred to as ‘reference scan’ below. A similar reference volume exists in BOLD scans as well. This allowed comparison of signal (change) levels between BOLD and SSPL data and aided in inter-scan image registration (see below).

Functional MRI used a 7.5 Hz full-field checkerboard visual stimulation task with 30 s ‘off’ / 30 s ‘on’ blocks, employing 5 ‘on’ blocks for the 110 repetition scans, and 10 ‘on’ blocks during the 210 repetition scans. The stimuli were presented with Presentation (Neurobehavioral Systems, Berkeley, CA, USA) via a PROPixx VPX-PRO-5050B projector (VPixx Technologies, Inc, Saint-Bruno, QC, Canada). Visual field of the images shown was approximately 7.6 × 5.8 degrees (w × h). Volunteer attention was monitored, and eye drift minimized, using a small center dot in the image that pseudo-randomly alternated between red and pink, events that the volunteer was expected to mark with a button press. A fORP 932 (Current Designs, Philadelphia, PA, USA) was used to record these button presses using the Presentation software.

2.3. Data analysis

All images were reconstructed and analyzed off-line using a combination of C and IDL (Harris Geospatial, Broomfield, CO, USA) based software. Parallel imaging reconstruction was performed using SENSE, as described previously [26].

Magnitude data were then spatially aligned using C-code based on software developed by Thévenaz [27] as follows: For each acquisition resolution, SSPL data were aligned to the last volume acquired. For SSPL, only in-plane rigid-body registration was used (2 translations and a rotation), since TI2-dependence over slices caused substantial contrast differences between slices (see Figure 1), hampering through-plane registration. The limited slice coverage would possibly also hamper 3D registration. The translation and rotation parameters found for the first SSPL volume were applied to the reference volume that had been acquired, without IR pulses, directly preceding it, assuming that no motion occurred during that brief interval. This was necessitated by the vastly different contrast between near-nulled SSPL data and non-inverted signal in the reference data. Subsequently, the BOLD data (both reference and functional volumes) were aligned to the SSPL reference volume using 3D rigid body registration (3 translations and 3 rotations), since these data all have similar image contrast. For FAIR scans, in-plane rigid body registration to the last image in the corresponding SSPL dataset was used for image registration.

Figure 1:

Results of SSPL pulse sequence optimization at 7 Tesla field strength using 2³ mm³ low-resolution experiments. (a) The reference scan for a slice, acquired without inversion pulses. (b) Data showing background signal suppression as a function of TI2 (in ms, yellow) for a 1400 ms TI1 for the imaging slice shown in (a). Images in (b) were scaled up 10x compared to the reference image in (a). In (b), the center slice of 5-slice low-resolution experiments (bottom row) is compared to the same slice from single-slice acquisitions using otherwise identical scan parameters (top row). Data show no notable detrimental effect of the 5-slice acquisition approach.

During initial inspection of the data, some aliasing artifacts were observed that had temporal frequencies of 1/12^th and 1/8^th of a Hz, in part attributed to inadvertent application of RF phase scrambling of the slice excitation pulses. The corresponding frequency bins were filtered out from the data. The resulting small loss of degrees of freedom (4 out of 110) was accounted for in the analysis.

Magnitude data were subsequently analyzed using a general linear model (GLM) that accounted for a hemodynamic response function (HRF) convolved paradigm. A gamma variate function with a half width and rise time of 3.5 s were used as HRF. The design matrix further included 2 polynomial trends, 3 (SSPL and FAIR) or 6 (BOLD) motion parameters, 4 respiratory and 4 cardiac phase regressors derived using an in-house python-based implementation of RETROICOR [28].

In addition to method-specific, t-threshold based masks, a so-called AND mask was defined, containing voxels that exceeded the t-threshold in both the BOLD and SSPL scan for that volunteer. The t-threshold used was ad hoc chosen as 4 for all data. This was a compromise between an overly cautious Bonferroni-corrected family-wise error threshold (which would be the threshold over which there is less than 5% chance that any one of the thousands of pixels in these data is a false positive) and the notably more lenient false discovery rate-based threshold, which would however vary depending on the size of the significantly activated area. This activation size dependence would complicate a fair comparison between BOLD and SSPL, especially at high resolution, where SSPL activation size is notably reduced compared to BOLD, as will be shown below. For the data in this work the appropriate thresholds for family-wise error were 4.84 ± 0.01 for the low-resolution scans and 5.10 ± 0.03 for the high-resolution scans, respectively. (Reported values are mean ± standard error over the thresholds for BOLD and SSPL combined at each resolution.) The t-threshold of 4.0 corresponds to p < 6.31E-5 and p < 4.51E-5 (uncorrected) for the 110- and 210-volume scans, respectively. The false-discovery rate thresholds were 3.40 ± 0.07 and 3.47 ± 0.09 for low-resolution BOLD and SSPL, respectively. For high-resolution scans, false-discovery rate thresholds were 3.43 ± 0.05 for BOLD and 3.79 ± 0.16 for SSPL.

To demonstrate that this choice of threshold and AND mask did not strongly affect the findings, we repeated the same analysis with t-thresholds of 3.5 and 5, respectively, as well as for an AND mask that was computed after the BOLD and SSPL masks for t > 4 were equalized in size by discarding the lowest-t-score voxels from the larger of the two masks in each pair, for each slice independently. This is referred to as the ANDeq mask.

In order to assess differences in large draining vein contribution in SSPL compared to BOLD, a vein-weighted ROI was defined based on temporal standard deviation [29]. Here, the 15% of voxels significantly activated in the BOLD experiment that exhibited the largest temporal standard deviation was included in a so-called VEIN-mask. Voxels significantly activated in both BOLD and SSPL that are within VEIN mask are referred to as the AND-V mask.

Various metrics used to compare BOLD and SSPL performance, such as temporal signal-to-noise ratio (SNR), image SNR, contrast-to-noise ratio (CNR), activation amplitude and background signal, were computed relative to the reference volume in each experiment. Since identical image contrast is expected for the BOLD and SSPL reference volume, using this volume as the basis for computing metrics allows for straightforward comparison of those metrics between the two techniques. Temporal SNR is defined as the reference image divided by the temporal standard deviation in the residue after fit of the functional timeseries data, after correction for the number of degrees of freedom lost by band rejection filtering and GLM analysis. Image SNR is based on this same reference volume and the square root of the noise variance estimate computed as part of the SENSE [25] reconstruction, which was derived from a noise measurement without RF excitation acquired at the start of each scan.

Unless otherwise noted all values are reported as the mean ± standard error over volunteers.

3. RESULTS

3.1. SSPL optimization

For single-slice acquisition, optimal suppression of the signal of gray and white matter was achieved at a TI1 of 1400 ms and a TI2 in the range of 450–475 ms (Figure 1). This led to a total delay (TI1+TI2) of close to 2 s, providing adequate perfusion weighting. At low-resolution, slice excitation and EPI readout took close to 50 ms, therefore most 5-slice SSPL experiments were performed using 350, 400, 450, 500 and 550 ms TI2. Comparison of 5-slice versus conventional 1-slice SSPL on possible detrimental effects of the additional radiofrequency (RF) pulses used to excite neighboring slices showed negligible differences (Figure 1). The increased number of slices required somewhat increased IR1 thickness. No notable detrimental effect of this on perfusion sensitivity, such as reduced perfusion signal because of the somewhat increased distance noninverted blood has to flow to reach the imaging slice, was observed (results not shown). This was confirmed for various TI2 ranges by using FAIR-like experiments, in which no changes in baseline perfusion level were found (results not shown).

3.2. fMRI task compliance

Compliance with the fMRI task was analyzed as previously described [30]. Inspection of the button-box response data for the various fMRI runs indicated that out of 18 pairs of SSPL and BOLD data (10 low-resolution and 8 high resolution), 12 were performed with high accuracy, resulting in a button press within 2 s following a dot color change on average 93.1% of the time. For the remaining 4 low-resolution and 2 high-resolution scan sets, however, at least one scan showed response accuracy below 80%. Those pairs were excluded from further analysis.

3.3. Low-resolution fMRI

Robust activation was found in low resolution SSPL experiments for TI2 ≤ 450 ms (Figure 2 and Table 1a). For those TI2 values, activation size was comparable to that found in BOLD experiments (Figure 3a), whereas shrinkage was seen at longer TI2s. This is partly attributed to the fact that residual BOLD signal in the SSPL data will counteract the perfusion signal at longer TI2: At longer TI2, the sign of the background signal is no longer inverted but rather has crossed zero and turned positive. Signal from inflowing blood on the other hand, only subjected to a single inversion, is still negative, even at the longest TI2 times. Increased perfusion adds negative signal to the positive background signal, whereas BOLD adds positive signal. This polarity difference between the perfusion and background signals also causes the apparently negative activation at longer TI2.

Figure 2:

Example activation t-score maps for one of the low-resolution (2³ mm³) SSPL experiments (bottom row) and the corresponding slices from the BOLD experiment performed in the same scan session (top row), superimposed on the reference scan for the respective acquisitions.

Table 1:

Summary of findings (± standard error over volunteers) for the comparison of SSPL to BOLD for various TI2 values and a TI1 value of 1400 ms; activation amplitude, CNR ratio, temporal SNR, % background signal and image SNR values are for the AND mask

(a): Low-resolution data (23 mm3) using TE=17.5 ms (n=6) and t=4
TI2 [ms] SSPL	activation size ratio SSPL/BOLD	activation amplitude [|%|]		CNR ratio SSPL/BOLD	temporal SNR		voxels			% background signal		image SNR
		SSPL	BOLD		SSPL	BOLD	AND mask	SSPL	BOLD	SSPL	BOLD	SSPL	BOLD
350	0.90 ± 0.19	1.11 ± 0.15	4.11 ± 0.62	0.91 ± 0.09	241 ± 28	97 ± 14	110 ± 17	200 ± 37	241 ± 30	7.5 ± 0.3	84.6 ± 1.5	453 ± 66	392 ± 57
400	0.97 ± 0.20	0.94 ± 0.14	4.41 ± 1.08	0.99 ± 0.10	274 ± 40	100 ± 17	93 ± 14	189 ± 32	214 ± 29	5.2 ± 0.2	85.0 ± 1.5	468 ± 71	411 ± 66
450	1.01 ± 0.23	0.81 ± 0.12	3.94 ± 0.86	0.91 ± 0.12	332 ± 47	107 ± 15	85 ± 15	178 ± 35	198 ± 30	3.1 ± 0.1	85.6 ± 1.3	508 ± 72	433 ± 61
500	0.57 ± 0.15	0.56 ± 0.07	3.55 ± 0.59	0.55 ± 0.12	392 ± 46	114 ± 18	51 ± 12	100 ± 22	192 ± 32	1.6 ± 0.1	86.4 ± 1.6	530 ± 63	448 ± 54
550	0.68 ± 0.18	0.50 ± 0.07	4.16 ± 0.60	0.54 ± 0.14	392 ± 55	98 ± 13	42 ± 13	94 ± 22	170 ± 42	2.4 ± 0.2	85.7 ± 1.6	546 ± 64	462 ± 55
(b): High-resolution data (13 mm3) using TE=38 ms (n=6) and t=4
TI2 [ms] SSPL	activation size ratio SSPL/BOLD	activation amplitude [|%|]		CNR ratio SSPL/BOLD	temporal SNR		voxels			% background signal		image SNR
		SSPL	BOLD		SSPL	BOLD	AND mask	SSPL	BOLD	SSPL	BOLD	SSPL	BOLD
180	0.51 ± 0.09	3.12 ± 0.33	8.04 ± 1.27	0.75 ± 0.08	61 ± 6	37 ± 3	170 ± 41	325 ± 65	613 ± 53	17.3 ± 0.9	85.2 ± 0.6	60 ± 6	55 ± 4
270	0.51 ± 0.13	2.42 ± 0.14	7.33 ± 0.72	0.68 ± 0.09	65 ± 3	39 ± 3	187 ± 62	333 ± 88	615 ± 73	11.8 ± 0.9	84.7 ± 2.0	64 ± 4	60 ± 3
360	0.34 ± 0.08	2.04 ± 0.24	7.08 ± 0.68	0.60 ± 0.14	78 ± 5	41 ± 5	216 ± 57	267 ± 78	665± 110	8.0 ± 0.6	83.2 ± 1.6	75 ± 4	71 ± 4
450	0.19 ± 0.05	1.63 ± 0.15	6.51 ± 1.17	0.58 ± 0.08	80 ± 4	41 ± 3	64 ± 18	118 ± 30	612 ± 101	4.6 ± 0.2	85.8 ± 1.4	67 ± 5	64 ± 5
540	0.08 ± 0.01	1.67 ± 0.33	9.36 ± 1.41	0.51 ± 0.08	98 ± 10	36 ± 5	10 ± 3	40 ± 9	536±103	4.0 ± 0.5	83.9 ± 2.8	88 ± 10	84 ± 11

Figure 3:

Plots of the various results shown in Table 1 as a function of the TI2 value used in SSPL. Recall that data from various TI2 delays are derived from different slices in SSPL; BOLD data for the corresponding slices are plotted on this same axis. Error bars shown are standard error over experiments (n=6). Plots show the number of significantly activated voxels (a); the activation amplitude as a percentage of the reference scan (b); the background signal level (c); contrast-to-noise ratio of the activation signal (d); image SNR (e); and temporal SNR (f).

Overall, the activation observed in SSPL was only minimally reduced compared to BOLD, both when comparing activation size and CNR (Table 1a and Figure 3a,d). This can be mostly explained by the higher temporal stability of SSPL (Figure 3f), which increased on the order of 3-fold when compared to BOLD, a likely result of reduced contribution of physiological noise due to the strongly suppressed background signal. Of note, differences in activation size and/or overlap between BOLD and SSPL could in part result from their different weighting towards arterial versus venous vasculature, and was documented previously [31].

Results further show that background suppression in SSPL is quite acceptable for all 5 slices in low-resolution experiments (Figure 3c). When compared to the reference volume without inversion pulses, the residual background signal is on average in the worst case 7.5%, at 350 ms TI2. Note that the reference volume is the first volume acquired in each scan, before there is steady state of the signal. It therefore has a slightly elevated signal level. This can be seen in the BOLD background signal level results, which are on average 85.5% of that scan’s reference volume (Figure 3c). In BOLD, no other contrast difference with its reference volume is expected other than a signal steady-state effect. Correcting for this steady state effect would yield an SSPL background signal level of 8.7% when compared to non-IR scans for low-resolution data acquired with 350 ms TI2. Even though the activation-induced signal change in BOLD is substantially larger than in SSPL (Table 1a), the signal change observed in SSPL is therefore still dominated by perfusion-related signal change for all 5 TI2 values employed here. Comparison with FAIR-like quantitative perfusion imaging (e.g. see Supplementary Material, Figure S2) showed that activation induced signal change in SSPL is also comparable to the signal change observed in FAIR-like fMRI techniques. Specifically, based on GLM analysis of the 4 datasets in the SSPL-FAIR comparison, activation-induced signal change as a percentage of the reference volume was 1.11% ± 0.12% (mean ± SD over volunteers) for SSPL and 0.87% ± 0.15% for FAIR. Baseline perfusion signal in these FAIR experiments was measured to be 0.76% ± 0.26%, meaning that the activation induced signal change as a fraction of the baseline perfusion signal was 124% ± 45% in those FAIR experiments.

A comparison of the mean signal time course for the AND-mask voxels in the slice for which TI2 is 450 ms in the SSPL experiment is shown in Figure 4a. The shaded area shows the standard error over volunteers. Both signals are shown relative to the signal level in the reference volume, so the y-axis used for each curve is different. These data further illustrate the similar performance of SSPL and BOLD at low resolution.

Figure 4:

Volunteer-averaged time-courses for BOLD and SSPL at low (a) and high resolution (b). Individual voxel time-courses in the AND mask in the slice for which SSPL had 450 ms TI2 were scaled to the corresponding signal in the reference volume before averaging. Note that different y-axes are used for BOLD (right) and SSPL (left). Shaded areas show the standard error over volunteers. The HRF-convolved block paradigm task regressor, scaled arbitrarily, is shown in grey in each pane.

Table 2a shows that the effect of t-threshold choice, or how the AND mask is derived, has little to no effect on the findings. As expected, mean activation amplitude increases and mask size decreases with increased t-threshold, but all TI2-dependent trends are the same.

Table 2:

Summary of findings (± standard error over volunteers) for the comparison of SSPL to BOLD for various TI2 values and a TI1 value of 1400 ms;

(a): Low-resolution data (23 mm3) using TE=17.5 ms (n=6) using TE=38 ms (n=6)						(b): High-resolution data (13 mm3)
SSPL TI2 [ms]	Activation size ratio SSPL/BOLD					SSPL TI2 [ms]	Activation size ratio SSPL/BOLD
	t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask			t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask
350	0.92 ± 0.19	0.90 ± 0.19	0.86 ± 0.19	1.00 ± 0.00		180	0.55 ± 0.08	0.51 ± 0.09	0.49 ± 0.12	1.00 ± 0.00
400	0.98 ± 0.19	0.97 ± 0.20	0.96 ± 0.22	1.00 ± 0.00		270	0.54 ± 0.12	0.51 ± 0.13	0.45 ± 0.13	1.00 ± 0.00
450	1.00 ± 0.22	1.01 ± 0.23	1.10 ± 0.29	1.00 ± 0.00		360	0.38 ± 0.07	0.34 ± 0.08	0.29 ± 0.08	1.00 ± 0.00
500	0.60 ± 0.14	0.57 ± 0.15	0.52 ± 0.17	1.00 ± 0.00		450	0.23 ± 0.05	0.19 ± 0.05	0.13 ± 0.05	1.00 ± 0.00
550	0.68 ± 0.17	0.68 ± 0.18	0.71 ± 0.19	1.00 ± 0.00		540	0.14 ± 0.03	0.08 ± 0.01	0.04 ± 0.01	1.00 ± 0.00
SSPL TI2 (*) [ms]	Activation amplitude BOLD [|%|]					SSPL TI2 (*) [ms]	Activation amplitude BOLD [|%|]
	t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask			t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask
350	4.00 ± 0.60	4.11 ± 0.62	4.61 ± 0.75	4.23 ± 0.82		180	7.38 ± 1.00	8.04 ± 1.27	8.69 ± 1.34	7.24 ± 1.60
400	4.14 ± 0.94	4.41 ± 1.08	4.80 ± 1.18	4.39 ± 1.05		270	7.02 ± 0.63	7.33 ± 0.72	7.59 ± 0.96	7.20 ± 1.37
450	3.72 ± 0.77	3.94 ± 0.86	4.22 ± 0.91	3.89 ± 0.84		360	6.35 ± 0.72	7.08 ± 0.68	7.51 ± 0.79	7.55 ± 0.69
500	3.52 ± 0.63	3.55 ± 0.59	3.52 ± 0.57	3.03 ± 0.58		450	6.40 ± 1.13	6.51 ± 1.17	7.54 ± 1.69	7.46 ± 1.27
550	3.77 ± 0.52	4.16 ± 0.60	4.97 ± 1.24	3.61 ± 0.56		540	8.05 ± 1.38	9.36 ± 1.41	13.07 ± 4.55	16.16 ± 8.27
SSPL TI2 [ms]	Activation amplitude SSPL [|%|]					SSPL TI2 [ms]	Activation amplitude SSPL [|%|]
	t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask			t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask
350	1.07 ± 0.14	1.11 ± 0.15	1.22 ± 0.17	1.19 ± 0.21		180	2.90 ± 0.33	3.12 ± 0.33	3.47 ± 0.34	3.13 ± 0.40
400	0.89 ± 0.12	0.94 ± 0.14	1.04 ± 0.14	0.99 ± 0.13		270	2.30 ± 0.13	2.42 ± 0.14	2.58 ± 0.20	2.31 ± 0.27
450	0.76 ± 0.11	0.81 ± 0.12	0.88 ± 0.13	0.85 ± 0.12		360	1.91 ± 0.17	2.04 ± 0.24	2.23 ± 0.28	2.39 ± 0.17
500	0.53 ± 0.07	0.56 ± 0.07	0.63 ± 0.08	0.56 ± 0.07		450	1.53 ± 0.12	1.63 ± 0.15	1.89 ± 0.19	1.69 ± 0.18
550	0.48 ± 0.06	0.50 ± 0.07	0.55 ± 0.06	0.54 ± 0.04		540	1.47 ± 0.17	1.67 ± 0.33	2.50 ± 0.86	4.80 ± 0.34
SSPL TI2 [ms]	CNR ratio SSPL/BOLD					SSPL TI2 [ms]	CNR ratio SSPL/BOLD
	t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask			t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask
350	0.90 ± 0.09	0.91 ± 0.09	0.92 ± 0.09	0.99 ± 0.10		180	0.74 ± 0.08	0.75 ± 0.08	0.80 ± 0.07	0.76 ± 0.09
400	0.96 ± 0.10	0.99 ± 0.10	1.00 ± 0.09	1.05 ± 0.10		270	0.66 ± 0.10	0.68 ± 0.09	0.70 ± 0.12	0.72 ± 0.11
450	0.90 ± 0.13	0.91 ± 0.12	0.92 ± 0.11	0.96 ± 0.15		360	0.61 ± 0.10	0.60 ± 0.14	0.62 ± 0.14	0.53 ± 0.18
500	0.55 ± 0.12	0.55 ± 0.12	0.66 ± 0.11	0.68 ± 0.14		450	0.56 ± 0.09	0.58 ± 0.08	0.68 ± 0.10	0.57 ± 0.16
550	0.53 ± 0.16	0.54 ± 0.14	0.68 ± 0.13	0.59 ± 0.15		540	0.31 ± 0.07	0.51 ± 0.08	0.80 ± 0.21	1.01 ± 0.19
SSPL TI2 [ms]	Mask size [#voxels]					SSPL TI2 [ms]	Mask size [#voxels]
	t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask			t >= 3.5 AND mask	t >= 4.0 AND mask	t >= 5.0 AND mask	t >= 4.0 ANDeq mask
350	129 ± 18	110 ± 17	81 ± 15	102 ± 21		180	217 ± 47	170 ± 41	108 ± 35	75 ± 33
400	111 ± 13	93 ± 14	67 ± 12	71 ± 16		270	233 ± 70	187 ± 62	128 ± 50	98 ± 51
450	105 ± 15	85 ± 15	60 ± 13	63 ± 17		360	226 ± 69	216 ± 57	135 ± 39	82 ± 34
500	63 ± 14	51 ± 12	30 ± 9	32 ± 11		450	89 ± 25	64 ± 18	34 ± 9	18 ± 8
550	54 ± 13	42 ± 13	29 ± 10	36 ± 12		540	19 ± 5	10 ± 3	3 ± 1	1 ± 0

^(*)BOLD voxels are separated by the TI2 in the corresponding SSPL experiment

Various signal characteristics in VEIN mask were investigated for SSPL at optimal (450 ms) TI2 and compared to BOLD for this same slice. Overlap between SSPL and BOLD activation was not significantly reduced in VEIN mask compared to the brain overall (−20% ± 22%). In SSPL, activation-induced percentage signal change relative to the mean brain signal was unchanged in VEIN mask when compared to the remainder of the BOLD-activated area (0.54% ± 0.21% versus 0.55% ± 0.12%). On the other hand, activation amplitude doubled for BOLD for this same comparison (6.96% ± 2.07% versus 3.11% ± 0.45%). Furthermore, in AND-V mask in BOLD data the temporal signal variance was found to be 171% ± 60% increased compared to AND mask overall, whereas it was not significantly different for SSPL at 450 ms TI2 (11% ± 20%) when comparing AND-V mask to AND mask.

3.4. High-resolution fMRI

Figure 5 shows sample SSPL data at 1³ mm³ resolution for optimal 450 ms TI2. When compared to similar low-resolution data in Figure 1, the effect of reduced image SNR can be clearly seen in an individual image (Figure 5a). However, when averaging data over the 5-minute experiment (100 repetitions, Figure 5b), it shows that the underlying image contrast is not affected, and that significant activation-induced signal change was present and detectable in the occipital lobe (Figure 5c–d).

Figure 5:

Sample 1 mm SSPL data for a single slice with 450 ms TI2. A single perfusion-weighed image from the functional timeseries is shown in (a), reflecting the substantially reduced image SNR at this resolution (e.g., compare to Figure 1). However, when temporally averaging 100 volumes, it can be seen that the underlying contrast is preserved (b). The fitted amplitude of the paradigm regressor in the GLM analysis for this slice is shown in (c). A thresholded t-map (t > 4) derived from (c) is shown in color superimposed on the reference volume for this slice, acquired at the start of the scan, is shown in (d).

The fMRI results at 1 × 1 × 1 mm³ are largely in line with the low-resolution findings (Figure 6 and Table 1b). The longer EPI readout train of high-resolution acquisitions forced larger spacing of TI2 over slices; values of 180, 270, 360, 450 and 540 ms were used, respectively. They show adequate sensitivity for the detection of activation, albeit less robust than the equivalent BOLD experiments. Reduced sensitivity of high-resolution SSPL is obscured by the fact that for suboptimal, short TI2 values, BOLD contamination is likely a significant contributor to the activation observed in SSPL, as is detailed below. For TI2 values with more optimal background suppression, the activated area size drops to below 20% of the activation area size found in BOLD (Table 1b and Figure 3a), suggesting that the advantage of improved temporal stability with SSPL seen at low resolution is waning. SSPL CNR at the optimal 450-ms TI2 is 42% reduced compared to BOLD (Figure 3d). It would therefore require 2.9-fold longer scans to compensate for this sensitivity loss. Comparison of the temporal stability values of SSPL and BOLD (Table 1b and Figure 3f) confirms that SSPL stability is still higher than in BOLD, (by about a factor of 2), but insufficiently so to compensate for the 3.6-fold larger activation-induced signal change observed in BOLD. A more significant contribution of thermal rather than physiological noise to the temporal stability in many of the voxels explains this observation. This loss of SSPL sensitivity relative to BOLD is further illustrated in the mean signal time-courses for the AND voxels in the slice that had 450 ms TI2 in the SSPL experiment (Figure 4b).

Figure 6:

Activation t-score maps for the 6 high-resolution BOLD (top) and SSPL (bottom) experiments, superimposed on the reference scan data. Rows show the different experiments, columns 5 slices for each of those experiments. Only approximately the back half of the slices is shown. The 3^rd and 4^th rows of each pane were the result of 630 s duration scans, the other four scans were 330 s in duration.

The results presented in Table 1b also suggest that the level of BOLD contamination in high-resolution SSPL data at sub-optimal TI2 can be higher than for the low-resolution experiments, especially for the 180 ms TI2 data, which is far from the 450–475 ms range that yields optimal background suppression. At 180 ms TI2, background signal level is on average at 17.3% of that in the reference scan. This corresponds to a BOLD contribution of 20.5% when correcting for lack of steady state in the reference scan (taking steady state signal to be on average 84.6% of the reference signal, as derived from the high-resolution BOLD experiment). Since the observed BOLD activation amplitude, or activation-induced signal change in BOLD data, is around 2.5-fold higher (8.04% versus 3.12%, Figure 3b) than the signal change in SSPL in the AND mask in the slice for which TI2 was 180 ms in SSPL, this would suggest that approximately half (20.5*(8.04/3.12) = 53%) of the signal observed in SSPL for could be BOLD contamination when using that suboptimal TI2. This is consistent with the observed 1.9 times higher activation amplitude in SSPL at TI2=180 ms (6.8%) when compared to activation amplitude at TI2=450 ms (3.5%). BOLD contamination in SSPL at high resolution is reduced to, on average, about 1/3–1/5 at TI2 values in the range of 360–450 ms, when background suppression is near optimal.

Similar to findings at low resolution, at high resolution the effect of t-threshold choice and masking approach is also predictable and minor (Table 2b). The only inconsistencies found are for large TI2 values, where significant activation is virtually absent and masks therefore only contain few voxels.

VEIN mask findings at high-resolution were in line with those at low-resolution. For data with optimal, 450 ms TI2, the overlap between SSPL and BOLD was found to be somewhat reduced (−31% ± 22%) in the VEIN ROI when compared to the BOLD activated region overall. In SSPL, activation-induced percentage signal change relative to the mean brain signal was similar in VEIN and the remainder of the BOLD-activation masks (0.43% ± 0.13% versus 0.46% ± 0.08%), whereas it was significantly increased for BOLD (8.98% ± 1.30% versus 5.70% ± 0.57%). The temporal signal variance was also notably increased in BOLD in AND-V (100% ± 55%), but not in SSPL data (9% ± 9%), when compared to all of AND mask.

4. DISCUSSION

To investigate the feasibility to perform single-shot perfusion-weighted fMRI at 1³ mm³ spatial resolution, we implemented a highly sensitive, single-short perfusion labeling technique at 7 T. Comparison of visual stimulation results with those of BOLD fMRI supported the notion that perfusion-based fMRI is relatively insensitive to the contribution of large veins. However, this came with a substantial loss in sensitivity in the detection of task-activation, primarily attributed to a lower activation-related absolute signal change. At the investigated scan time of about 5 minutes, this affected the ability to detect brain activity. This sensitivity reduction was not present at lower resolution, where the lower signal change was accompanied by a reduced level of physiological noise.

4.1. Low-resolution experiments

At low (2³ mm³) resolution, SSPL performance was on par with BOLD, demonstrating the importance of reducing physiological noise, which is well known to be a limiting factor in fMRI performance under conditions of high image SNR [26, 32]. The dual inversion preparation in SSPL suppressed in excess of 90% of the background signal for all TI2 values investigated, while minimally affecting the perfusion signal. As a result, temporal instabilities in the background signal, caused by effects such as head motion, respiration and other non-neuronal sources, were substantially reduced. Furthermore, as previously demonstrated for SSPL (Figure 8 in [17]), adequately suppressed stationary background signal in SSPL virtually eliminates BOLD contamination. This is in line with an earlier demonstration of the perfusion sensitivity of a selective inversion pulse, and the absence of a BOLD contribution, as demonstrated using non-selective inversion [22]. The absence of perceivable task-induced signal change in the average signal time course in FAIR control images (Supplementary Material Figure S2e) also illustrates this.

The reduction in temporal noise counter-balanced the lower activation amplitude of the perfusion signal relative to BOLD. Inherent to its perfusion-dominated contrasts, SSPL thus allows for a reduced spatial blurring from large veins, a prominent feature of BOLD fMRI. Comparison of BOLD and SSPL data in a vein-weighted mask indicates that the functional response in SSPL data, unlike BOLD data, is not notably affected by the presence of large veins. That said, signal from larger arteries was not explicitly suppressed in these SSPL experiments and could thus contribute to the task-induced signal change observed. The size of that effect could be investigated, and flow-suppressing gradients added if needed. Furthermore, SSPL, as implemented here, has reduced spatial coverage compared to BOLD, owing to the limited range of TI times at which background suppression is effective. It should also be mentioned that the TE value used for the low-resolution experiments was sub-optimal for BOLD, where the optimal sensitivity occurs at the ~ 30 ms T₂* of grey matter [33].

4.2. High-resolution experiments

The implementation of 1 mm resolution perfusion-based fMRI revealed several difficulties and potential disadvantages. In contrast with the 2 mm resolution results, the 1 mm resolution perfusion data showed that a substantial sensitivity loss occurred relative to BOLD. Whereas this could in part reflect reduced sensitivity to signal from large draining veins, the finding of reduced sensitivity is predominantly attributed to the diminished advantage of reduced physiological noise with background suppression in SSPL. Because of the greatly reduced image SNR in the high-resolution experiments (Figure 3e, and evident when comparing Figure 1 and Figure 5a), physiological noise (and advantages of its reduction) becomes overwhelmed by thermal noise and therefore less relevant. As a result, the smaller activation amplitude of the perfusion signal is accompanied by a lower noise level, and therefore directly translates in a reduced sensitivity and a loss in robustness. Thus, robust, high-resolution perfusion-based fMRI will likely require additional technical improvements, or substantially increased signal averaging by lengthening the scan time.

A couple of other issues affected practical implementation and performance of high resolution SSPL-based fMRI. Two of these issues are related to the lengthening of the EPI readout train with higher resolution scanning. Firstly, for the perfusion signal, the increased T₂* related signal loss reduces the signal of interest, while BOLD sensitivity (and thus potential contamination) increases. This may be remediated somewhat by using SSPL with a spin-echo acquisition or the use of higher in-plane image acceleration factors.

Secondly, the longer EPI readout duration also forces a larger spread of nominal TI2 times across slices, leading to larger deviations from optimal background signal suppression in some of the slices in these experiments. At the same time, at high resolution, more slices will be required to cover the brain area of interest. While some of this could be remediated by multi-band acquisition, as well as higher in-plane image acceleration, slice coverage will remain much inferior to BOLD and far below including the whole brain. Due to the long TR required in perfusion fMRI, and a similar need for signal acquisition close to optimal TI2, an efficient 3D implementation of SSPL is also not practically feasible without incurring a significant degree of blurring and/or BOLD contamination. As has been previously demonstrated [34], the latter could be reduced by a more conventional subtraction-based imaging scheme, albeit at the cost of increased sensitivity to physiological noise, noise amplification due to the subtraction, and reduced temporal resolution.

Thirdly, due to the combination of limited coverage in the slice direction combined with TI2-dependent contrast variation over slices in SSPL, motion correction had to be limited to in-plane only. The inclusion of 6 motion regressors in analysis of BOLD data led to a notably larger improvement than inclusion of 3 motion regressors in SSPL analysis (results not shown), illustrative of this drawback.

4.3. Previously published high-resolution perfusion-weighted fMRI in humans

SSPL is not a widely used perfusion fMRI technique, so to put the findings in this work into perspective we refer here to a limited, and therefore incomplete, number of previous high-resolution perfusion-based fMRI studies in humans to illustrate that a finding of reduced activation size compared to BOLD, and the need for substantial temporal averaging to boost sensitivity, are archetypal.

Early work at 7 T by Pfeuffer et al. used visual stimuli, down to 0.9 × 0.9 × 1.5 mm³ (1.2 μl) resolution. At that resolution, a 4-fold reduction of activation size was found for FAIR when compared to BOLD fMRI in 6-minute scans also employing a 3 s TR [35]. A more recent in-depth comparison of various ASL techniques at 3 and 7 T down to 1.5 × 1.5 × 3.0 mm³ was performed by Ivanov et al. [36]. Very recently, to our knowledge the first perfusion-imaging study showing laminar fMRI, based on data acquired at 0.9³ mm³ resolution, was published. In this work, as stated due to low SNR, 44 min of averaging was required to derive the laminar perfusion data [37].

5. CONCLUSIONS

The feasibility of high-resolution SSPL for perfusion-based fMRI at 7 Tesla field strength, at up to 1 × 1 × 1 mm³ (1 μl) nominal resolution and 3-s temporal resolution, was demonstrated in limited brain areas. Perfusion contrast appeared less affected by large veins, and therefore may have higher spatial fidelity to the underlying neuronal activity changes. However, at 1 mm isotropic resolution and with short (5½-minute) scan time, a reduction in sensitivity and robustness of around 42% was observed relative to BOLD fMRI, which was not apparent at lower (2 mm isotropic) resolution. It appears therefore that at resolutions below 1 mm, the sensitivity loss of perfusion-based fMRI may become too severe and outweigh its reduced sensitivity to venous blurring.