Combined prospective and retrospective motion correction for high-resolution linescan acquisitions in the human cerebral cortex at 7T

Introduction

Classical 1D linescan acquisitions^1–3 have recently been revived in several “laminar MRI” studies of diffusion⁴, relaxometry^5,6 and BOLD fMRI signals^7–15. The key idea is that, by prescribing each line perpendicular to the cortical surface at a location of interest, extremely high resolution can be achieved in the radial direction of cortex, raising the eventual possibility of recording signals from individual cortical layers.

However, a common concern with these and other reduced field-of-view acquisitions is the possibility of head movement during the scan. We therefore incorporated volumetric navigators^16,17 (vNavs) into our linescan pulse sequence and demonstrate here that combining prospective motion correction (PMC) with 1D retrospective motion correction (RMC) enables reliable linescan measurements at 7T with 0.5-mm readout resolution.

Methods

Two healthy volunteers (1F/1M, ages: 28–29 years), having provided written informed consent in accordance with our Institution’s Human Subjects Research Committee, were scanned on a Siemens Terra 7T whole-body scanner using an inhouse-built 64-channel head-receive and birdcage transmit coil¹⁸.

Reference transmit voltages were determined online using the NeuroSpin “ns_tfl_rfmap” C2P package and a 1×1×1 mm³ anatomical FOCI-MEMPRAGE scan¹⁹ was used to prescribe a line roughly perpendicular to primary sensorimotor cortex on opposite banks of the central sulcus.

Fig. 1 shows the diagram for our pulse sequence, where a 3D-EPI-based volumetric navigator^16,17 (vNav) follows each LSGESSE (Line Scan with Gradient-Echo Sampling of a Spin Echo) acquisition^5,15.

Fig. 1:

Pulse sequence diagram. LSGESSE slice-select gradients for excitation and refocusing pulses are played orthogonally to one another, resulting in signal from just a column of spins (inset). Thus only 1D spatial Fourier encoding is required, with high resolution possible in the readout direction. A single spin echo is sampled by multiple gradient echoes, from which transverse relaxation rates R₂=1/T₂ can be derived. A 3D-EPI vNav is played after an interval τ, followed by a delay time t_d during which vNav images are reconstructed and the motion update is calculated and applied.

LSGESSE parameters were as follows: excitation/refocus FA = 90°/180°, 7 unipolar gradient echoes with the 4th gradient echo coinciding with the spin echo at TE = 57 ms (BW = 257 Hz/px); the voxel size along the line was 0.5 mm, with a 256-mm readout field-of-view and nominal line thickness of 3 mm.

For the vNavs: FA_vNav = 2°, TR_vNav/TE_vNav = 5/2 ms, BW_vNav = 5580 Hz/px (nominal echo spacing = 0.28 ms), voxel size = 7.5×7.5×7.5 mm³ (matrix = 32×32, partitions = 24) and GRAPPA factor = 5×1.

For each linescan “run”, we acquired 30 repetitions (excluding dummy scans) with a TR of 2 s (t_LSGESSE/τ/t_vNav/t_d = 100/1650/125/125 ms; see Fig. 1).

Each subject was asked to remain as still as possible from the start of the session to the end of the first block of linescan runs. For each run in the second block, the subject was instructed (via intercom) to move their head to a new position at the 30-s mark, hold that position until the end of the run (60-s mark), and then to try their best to return to the original position.

Results

We first sought to assess, in the absence of motion, the effect that incorporating vNavs and applying PMC might have on linescan data quality. Fig. 2 shows linescan data acquired with vNavs+PMC versus without vNavs, for the first versus the last 25s of the run. The overlap of the curves indicates that (1) the subjects indeed appear to have stayed still during these runs (as instructed) and (2) any difference in data quality is too subtle to be appreciated by eye.

Fig. 2:

In the absence of head motion, vNav-based PMC appears to do no harm. Yellow lines and circle indicate line orientation and center, overlaid on the MPRAGE data (top). The region with solid yellow lines corresponds to the zoom-in of the middle and bottom panels, where the LSGESSE ECHO 4 signal intensities and corresponding R₂ values, respectively, are plotted versus distance along the line, averaged over the first versus last 25s of the run, and for no vNavs versus vNavs+PMC. The slightly lower overlap of the R₂ curves may be due to noise amplification inherent to fitting exponentials.

In the presence of motion, Fig. 3 shows that RMC alone (i.e., without PMC) fails to be adequate, whereas Fig. 4 shows that the combination of PMC and RMC provides excellent results.

Fig. 3:

In the presence of head motion, RMC alone does not suffice. The LSGESSE ECHO 4 signal intensity (top) and R₂ (bottom) curves from the last 25s (solid blue), after the head movement, have a fundamentally different shape to those from the first 25s (red), and therefore even our best attempts at applying shifts for RMC (dashed blue) fail to adequately align these curves. Note that (for the first 25s) the acquired lines may no longer be exactly in the position shown in the top panel of Fig. 2, depending on the subjects’ accuracy in returning to the original head position at the end of each run.

Fig. 4:

In the presence of head motion, the combination of PMC and RMC appears to work best. With vNav-based PMC enabled, the LSGESSE ECHO 4 signal intensity (top) and R₂ (bottom) curves from the last 25s (solid blue), after the head movement, and the first 25s (red) can be well aligned by applying shifts along the line for RMC (dashed blue). Note the slight difference in vertical scaling for the ECHO 4 intensities in Subject 1, likely due to small differences in the voxelwise receive-coil sensitivities for the two different head positions; such differences drop out in the R₂ estimates.

Discussion

A recent study investigated navigator-based PMC of linescan acquisitions¹³ but has several important differences to our work: (1) RMC was not considered; (2) saturation pulses preceding slice-selective excitation were used to define the line (i.e., outer-volume suppression); (3) linescans were acquired once every ~0.1 s; and (4) navigators of 0.5–1.7 s duration were acquired once every ~45 s (thus heavily disrupting the steady-state magnetization of the linescans), whereas we currently acquire one navigator per linescan (i.e., every TR). We note that these authors sought to acquire gradient-echo BOLD-fMRI data with extremely high temporal resolution, which may have dictated some of their choices, whereas our spin-echo-based linescans with a longer TR present a very different set of constraints (and opportunities).

Lowering the TR amplifies “burn-in” or spin-history artifacts in the vNav images, as shown in Fig. 5. Furthermore, the concurrent acquisition of multiple lines²⁰ may produce artifacts with more complicated spatial patterns. Understanding the effect of these more pronounced or complicated artifacts on the estimates of motion, and the role of advanced registration algorithms that could mask out these artifacts, will require further investigation.

Fig. 5:

Lowering the TR from 2 s (top) to 500 ms (bottom) leads to increased “burn-in” or spin-history artifacts in the vNav images, seen as a reduction in signal intensity at the location of the LSGESSE excitation and refocusing planes. These artifacts result from insufficient longitudinal magnetization recovery of the spins at these locations during the interval τ (see Fig. 1) and is exacerbated by the fact that the optimal excitation flip angle for short-TR spin-echo acquisitions exceeds 90° for typical T₁ values in the brain. For TR = 2 s, these artifacts do not appear to impact PMC much.

In conclusion, the preliminary results shown here are highly encouraging for single-line, long-TR, spin-echo-based applications^4–6,15. However, further testing and assessment of robustness is in order, especially given a greater variety of head movements and associated changes in B₀ inhomogeneity.

Synopsis.

Motivation:

Classical 1D linescan acquisitions have recently been shown to be valuable for recording in-vivo MRI signals across the layers of human cerebral cortex; however, these reduced-field-of-view techniques are vulnerable to rotational head movements as well as translations parallel and perpendicular to the line.

Goal(s):

To acquire high-quality, high-resolution linescan data that is robust to “in-line” and “through-line” motion.

Approach:

3D-EPI volumetric navigators (vNavs) were incorporated into a spin-echo-based linescan pulse sequence.

Results:

We demonstrate that, by combining prospective and retrospective motion correction, we can acquire reliable linescan data with 0.5-mm readout resolution at 7T, in the presence of in-line and through-line head motion.

Impact:

Motion-robust linescan techniques will help enable the measurement of tissue microstructure and microvascular fMRI signals at high spatial resolutions, approaching the thickness of individual cortical layers, facilitating noninvasive studies of cortical circuitry and architectonics in the living human brain.