Arterial spin labeling (ASL) is a perfusion-based MRI technique with great potential to study neurodegenerative diseases such as frontotemporal dementia (FTD). One obstacle in multi-centre studies is the variability attributed to differences in ASL labeling and readout strategies, site hardware and scanner type ‒ i.e., vendor^{1, 2}. These sources affect cerebral blood flow (CBF) images through differences in hemodynamic contrast, signal to noise ratio (SNR), spatial smoothing and geometric distortion³. Inter-vendor and inter-site variability may significantly degrade the statistical power to detect meaningful perfusion abnormalities at the level of group inference⁴. In the present study we compare several spatial registration methods to reduce this variability.

Data were drawn from the GENetic Frontotemporal dementia Initiative (GENFI)⁵, a large multi-centre study that aims to identify the earliest brain changes in individuals who have a genetic risk of developing FTD. The multi-modal MRI protocol included ASL with the caveat that GENFI included four different ASL implementations by necessity (Table 1). 12 GENFI control subjects ‒ i.e., mutation-negative ‒ were randomly selected for each of the ASL implementations ‒ i.e., 48 subjects in total. 3D volumetric T1-weighted scans (T1) of each subject were segmented, masked and registered to MNI space using SPM12 and DARTEL6. CBF-maps were computed and rigidly registered to proton density (M0) ASL reference scans⁷. Two approaches were evaluated, a T1-approach in which masked M0 images were registered to masked native space 3D-T1 images and a pGM-approach in which masked CBF images were registered to gray matter probability maps (pGM, obtained via SPM12 segmentation). A multi-step masking algorithm was used, involving a dilated MNI brain mask registered to a thresholded CBF map. The same brain mask was used for both approaches. Consecutive SPM12 registration options that were evaluated included: 1) rigid only - i.e., 6 parameter rigid-body, "coregister"; 2) an additional uniform non-linear registration - i.e., 12 parameter affine + elastic, "old normalize"; for the pGM-approach also 3) an additional local adaptive non-linear registration - i.e., creating a DARTEL template for each subject from a CBF and a pGM image^{6, 8}. To avoid multiple interpolation errors, all transformations were combined before resampling CBF images into a 1.5 mm isotropic resolution common space. The resultant CBF images were scaled to a mean GM CBF of 50 mL/100g/min per ASL implementation.

To quantify the overlap between two scans, the mean voxel-wise Generalized Tanimoto Coefficient was calculated within a MNI GM mask⁹. This overlap measure can be interpreted as a more strict Dice Similarity Coefficient or Kappa statistic⁹ and will be referred to as similarity coefficient (SC), ranging from completely dissimilar (0) to identical images (1). Assuming that perfect registration should not lead to identical images but still retain physiological CBF differences, SC > 0.7 can be regarded as excellent agreement⁹. The SC was computed for each pair-wise image comparison, resulting in (n[n-1])/2 = 66 and 1128 unique comparisons for n=12 and n=48, respectively.

Performance of the registration procedures was evaluated visually, i.e., through improved tissue contrast on the mean CBF map (Figure 1). Figure 2 shows mean histograms of the between-subject CBF similarity. Across all ASL implementations, the worst registration was observed for the uniform non-linear approaches (Figure 1.2 and 1.4 and black lines Figure 2.1-2). The rigid pGM-approach visually outperformed the rigid T1-approach for 3D GRASE (Figure 1.3d vs. 1.1d and red lines Figure 2.d) and on the combined multi-centre images (Figure 1.3e vs. 1.1e and red lines Figure 2.e) but not the other ASL implementations (Figure 1.3a-c vs. 1.1a-c and red lines in Figure 2.a-c). The local adaptive non-linear pGM-approach improved registration for 2D EPI Bsup only (Figure 1.5b and green line Figure 2.2b). There were significant SC differences between the five registration methods (p<0.001, FWE-corrected Kruskal-Wallis test, Table 2).The multi-centre CBF images produced in the current work show that the choice of image registration approach has a profound effect on ASL studies that involve multiple pulse sequences and/or sites. By inverting the transformations, it can be expected that these comparisons in standard space are equally valid for registration of MNI-atlases via anatomical images into subject ASL space, although this may depend on the ASL acquisition resolution. The optimal similarity that could be achieved with local adaptive non-linear registration for 2D EPI Bsup is promising, but requires further research investigating its effects on partial volume fractions^{8, 10}. Currently, the rigid-body registration of the CBF image to the pGM image seems to produce the most robust registration as a standard approach across the different ASL implementations.