Introduction

Previously (ISMRM 2020¹), we highlighted an issue with conventional volume registration techniques used to correct motion in cerebrovascular function studies that induce global intensity changes (GICs) using CO₂ stimuli and ASL tagging. We determined that these volume registration algorithms (VRA’s) evaluate GIC’s at edge voxels to be motion, thus causing relevant signal loss when motion parameters are regressed from the data. Here, we compare VRA-based and an external marker-less camera-based correction with novel techniques designed to limit the influence GICs on the estimated motion parameters.

Methods

Image Acquisition
Data were collected on a Siemens 3T MAGNETOM Prisma scanner with a 64-channel receiver head coil. The scanning protocol was:
· Multi-PLD pCASL (TR1,2=5.6/3.1s,TE=11ms, TI=250-3000ms in 250ms steps, α=90°,In-plane=3.4mm², 20 slices(6mm)). Collected to quantify CBF. Expected TI ramp and ASL tagging to drive errors in VRA motion parameter estimate.
· Dual-excitation pCASL (TR=1.1/3.8s,TE=11/30ms, α=90°, In-plane=3.4mm², 19 slices(6mm)). Collected to quantify CVR, CBF, OEF and CMRO₂. Participants underwent periods of hypercapnia and hyperoxia. Expected CO₂ increases to drive erroneous motion estimate in the BOLD portion and a mixture of ASL tagging and CO₂ to drive error in ASL motion estimate.
· Breath Hold (TR=2s,TE=30ms,α=70°,In-plane=2mm²,64 slices(2mm)),TR=2s,TE=30ms, In-plane=2mm² ,64 slices(2mm)). Collected to quantify CVR. Expected CO₂ to drive errors in motion estimate although expected large amounts of motion to be present also.
· Resting-state (TR=2s,TE=30ms,α=70°,In-plane=2mm², 64 slices(2mm)). Collected as a control without any GIC, connectivity measures were quantified. Expected motion estimates to perform in a similar way across all motion-correction methods.
· Resting-state with applied motion. The protocol was identical to the Resting-state. Participants were asked to move their head along with a predefined movement sequence. Collected to determine how well methods perform in severe motion conditions, connectivity values were quantified. Expected similar results to that of resting-state.

Camera Motion collection
The marker-less camera (TracInnovations, Denmark) acquired motion data throughout the scan by using infra-red light to create a ‘mesh’ on the face to track motion. This data was calibrated to the scanner coordinates using an MPRAGE to create final motion parameters.

Data Analysis
Estimates of motion were calculated in four groups of methods; VRA-based, camera-based, ICA-based, and Erosion-based. The VRA-based method used AFNI’s 3dAllineate^2,3 to perform rigid-body volume registration and calculate motion parameters. The camera-based method estimated motion parameters using a marker-less system. The ICA-based method used an Independent Component Analysis⁴ and the theoretical signal expected to be caused by any CO₂ or tagging induced global changes, to isolate and remove components related to these theoretical signals. Then a VRA was used to calculate motion parameters on the dataset stripped of the components containing these theoretical signals. The severity of removal was based on a correlation of the theoretical signal with each ICA component. In method 1 the ICA components that had p-values of statistical significance at p<0.05 (uncorrected) were removed, whereas in method 2 a Bonferroni correction accounted for multiple comparisons and statistically significant components were removed. The erosion-based method excluded voxels from the VRA using a whole brain mask which had been eroded (from the outer edge) by either 5mm or 20mm.
All analyses used linear regression with the motion parameters (from each correction type) as regressors from either the global signal (Fig 1), the theoretical GIC, hypothesised to drive the error in the VRA (Fig 2) or the VRA motion parameters (Fig 3). R² values were calculated from each regression and paired t-tests were used to estimate the significance when compared to the VRA's R² values.

Results

In Figure 1, the erosion methods showed R² values consistent with that of the VRA estimate for most scans. The ICA methods displayed a lower R² value than that of the VRA in most cases suggesting less signal is lost. For the Resting + CO₂ scan, p-values of 2.2x10^-4 and 1.6x10^-4 were calculated for ICA method 1 and 2 respectively. In Figure 2, the ICA methods showed a lower R² in all scan types with p-values of 2.1x10^-4 and 5.8x10^-3 (Resting + CO₂) for method 1 and 2 respectively. The ICA methods showed lower R² values in the Resting-state scan with p-values of 7.8x10^-3 and 2.6x10^-3 for method 1 and 2 respectively. In Figure 3, the erosion value of 20mm showed low values of R² for some scans.

Discussion

We’ve shown that the use of an ICA to remove signal pertaining to the GIC’s can lead to more accurate motion estimates containing less information concerning these GIC’s. However, in some cases this may not be enough and other methods should be incorporated to remove the GIC prior to correcting for motion, for example, in multi-PLD ASL scans where a TI ramp drives a global signal increase. The use of an eroded mask can in some cases aid this estimation of motion but in many scenarios, this can lead to erroneous results similar to VRA estimates. In the case of scans where no GIC’s are present one could argue that the VRA method is sufficient, however given that breath holding results in an increase in CO₂, this could affect the reliability of VRA estimates and as such the ICA methods could aid in this estimation.

Acknowledgements

This work was supported by Wellcome [WT200804]