Overview of the UK biobank ASL data and processing pipeline.

The raw calibration (M0) and label/control images at each postlabelling delay (PLD) are pre-processed to correct for motion, B0-induced distortion and gradient non-linearities and the ASL data are control-label subtracted to give corrected M0 and perfusion-weighted images (PWIs) at each PLD. Kinetic model fitting in each voxel and calibration with the M0 image gives an estimate of absolute cerebral blood flow (CBF) and arterial transit time (ATT) in each voxel. The ASL data are registered to the structural T1-weighted images processed by the original UK biobank pipeline, which include the necessary warp fields from standard (MNI) space, allowing tissue-type masks and regions of interest to be overlaid on the ASL maps and regional mean values to be extracted as Image-Derived Phenotypes (IDPs). The CBF and ATT maps from each subject are also registered to standard space to allow voxelwise cross-subject analyses to be performed and group average templates to be derived.

Simple associations and deconfounding.

Despite the short scan time available for each subject, the mean grey matter (GM) CBF shows the expected robust decrease with age, along with a concomitant increase in ATT (shown as a sliding window mean and standard deviation in A, window width 5 years). This robust correlation is also present at the voxelwise level (B), with most of the brain showing strong correlations of CBF and ATT with age (shown here as maps of the Pearson correlation coefficient, r, in standard space). Small variations across the day (C, sliding window width 2.6 hours) and with sex (D) are also evident. The effect of haematocrit on the ASL IDPs is significant (E), but is removed by the deconfounding process.

Arterial transit times and postlabelling delay considerations.

A) The cumulative distribution of arterial transit times (ATTs) across all grey matter voxels in all subjects shows that more than 98% of voxels have ATTs of 2s or less in this cohort. B) These distributions can also be used to calculate the 95th or 99th percentile ATT values in grey matter for each subject. The 95th or 99th percentiles of these values in a sliding window across ages (window width 5 years) is plotted here, suggesting the postlabelling delay (PLD) necessary to accurately estimate CBF with single-PLD ASL in this proportion of voxels and subjects.

Correlations between ASL and other (non-imaging) measures.

A) Manhattan plot showing the significance of univariate associations between ASL IDPs and all the non-imaging variables grouped and colour-coded by category along the x-axis. Each dot represents the most significant association with an ASL IDP, plotted as −log10(P) on the y-axis, with larger values being more significant. Bonferroni and false discovery rate (FDR) multiple comparison correction thresholds are also shown. Some interesting associations exceeding the multiple comparisons thresholds are highlighted. B) Spatial correlation maps corresponding to selected significant associations with non-imaging variables (shown as Pearson’s correlation coefficient maps, r, in standard space with positive correlations in red-yellow and negative in blue-light blue, overlaid on the MNI template using radiological convention L=R). Some of these factors have a near-global effect on CBF or ATT, while others are more spatially specific, such as the lower CBF in frontal regions associated with alcohol intake, ATT increases in deep grey matter with hypertension and CBF reductions particularly in right opercular cortex associated with poorer performance in a cognitive task (numeric path duration).

The sensitivity of ASL to other (non-imaging) measures.

A) Manhattan plot showing the strength of associations between non-imaging measures and the ASL IDP group relative to similarly sized groups of IDPs from other modalities obtained from permutation testing (see Methods for details). −log10(P) values exceeding the FDR threshold indicate that ASL is significantly more strongly associated with this non-imaging measure than other modalities. Note that the maximum −log10(P) value achievable is limited by the number of IDP groups and permutations used. All non-imaging measures were included in statistical testing, but only those which were significantly correlated with at least one ASL IDP are shown here for clarity. B) Two examples of health-related measures where ASL was more sensitive than other modalities. Subjects who have suffered from a depressive episode had significantly longer ATTs, particularly in the left frontal pole, as well as sub-threshold CBF reductions in a number of other brain regions (see Supplementary Figure 2). Hearing difficulty was significantly associated with reduced CBF in several auditory processing regions including right Heschl’s gyrus, precentral cortex and frontal regions, as well as widespread ATT increases. To show these subtle associations more clearly, the statistically significant regions are highlighted with a black outline and overlaid on the spatial correlation maps, with opacity modulated by the correlation coefficient, as recommended by (Taylor et al., 2023).

Correlations between ASL and other brain imaging measures.

A) Manhattan plot of the univariate associations between ASL IDPs and all other brain imaging IDPs, grouped and coloured by modality along the x-axis. Many associations exceed the Bonferroni and FDR thresholds for significance, with a few non-ASL IDPs of interest highlighted here. B) Voxelwise spatial maps of correlation coefficients between CBF/ATT and some selected non-ASL IDPs. These include increased CBF in regions with higher white matter/grey matter contrast, reduced CBF and increased ATT where more white matter hyperintensities are found, associations with deep grey matter T1-weighted intensity and T2*, and even increased CBF and reduced ATT in white matter regions associated with higher fractional anisotropy (FA, derived from diffusion tensor imaging).

Associations between ASL and resting fMRI.

Correlation coefficient maps of CBF against selected node amplitudes are shown in order of decreasing significance (top to bottom and left to right). The regions associated with each node are highlighted with a white outline (z > 5). It appears that higher node amplitudes generally correspond to higher CBF within the node region, or reduced CBF outside that region.

Examples of spatial maps where correlations with ASL IDPs may be artefactual.

The smoothly varying CBF reductions at the brain edges that are associated with weight and bone mineral density could be driven by differences in magnetic field distortions or the ability of the registration algorithms to tackle variations in these physical factors. CSF volume likely drives partial volume effects, leading to apparent reductions in CBF adjacent to regions containing CSF. The optic chiasm is adjacent to brain feeding arteries, so this association is likely driven by artery morphology. Similarly, R Ant Amygdaloid Area AAA is a measure made adjacent to the right posterior cerebral artery, so estimation of this area is likely correlated with the artery diameter, leading to an apparent association in the vascular territory of this artery.

Additional spatial maps demonstrating the association between CBF and depressive episode.

Whilst there are no significant clusters, regions showing higher correlations have previously been found to have altered connectivity in depression (Helm et al., 2018).