Rapid processing and quantitative evaluation of structural brain scans for adaptive multimodal imaging

Abstract Current neuroimaging acquisition and processing approaches tend to be optimised for quality rather than speed. However, rapid acquisition and processing of neuroimaging data can lead to novel neuroimaging paradigms, such as adaptive acquisition, where rapidly processed data is used to inform subsequent image acquisition steps. Here we first evaluate the impact of several processing steps on the processing time and quality of registration of manually labelled T1‐weighted MRI scans. Subsequently, we apply the selected rapid processing pipeline both to rapidly acquired multicontrast EPImix scans of 95 participants (which include T1‐FLAIR, T2, T2*, T2‐FLAIR, DWI and ADC contrasts, acquired in ~1 min), as well as to slower, more standard single‐contrast T1‐weighted scans of a subset of 66 participants. We quantify the correspondence between EPImix T1‐FLAIR and single‐contrast T1‐weighted scans, using correlations between voxels and regions of interest across participants, measures of within‐ and between‐participant identifiability as well as regional structural covariance networks. Furthermore, we explore the use of EPImix for the rapid construction of morphometric similarity networks. Finally, we quantify the reliability of EPImix‐derived data using test–retest scans of 10 participants. Our results demonstrate that quantitative information can be derived from a neuroimaging scan acquired and processed within minutes, which could further be used to implement adaptive multimodal imaging and tailor neuroimaging examinations to individual patients.

CSF is low or nulled -in the case of T 1 -FLAIR, an inversion recovery sequence is used in which the inversion time is chosen such that the magnetisation from CSF passes through zero. In the IR-FSPGR T1-weighted sequence, CSF also has low signal. In this case, however, this is not due to choice of inversion time, but due to the relationship between contrast, flip angle and repetition time. Importantly, we refer to both contrasts as simply "T 1 -weighted", or "T 1 -w".

Spherical permutations of ROIs
Spherical permutations were generated by randomly rotating a projection of ROI centroids on the (FreeSurfer) sphere, before mapping rotated ROIs to the nearest unrotated ones. Mirrored rotations were applied to the contralateral hemisphere, resulting in a permutation which controls for spatial autocorrelation and hemispheric symmetry of regions (Váša et al., 2018;Alexander-Bloch et al., 2018;Markello and Misic, 2020). P-values for the correlation between two regional maps were obtained by comparing the empirical value of Spearman's ρ to a null distribution of Spearman correlations, generated by correlating one of the empirical maps to a set of 10,000 spatially permuted versions of the other map; these "spin-test" P-values are referred to as P spin . Spin-test P-values were additionally corrected for multiple comparisons using the false discovery rate (FDR; Benjamini and Hochberg, 1995).

Intrinsic connectivity networks
We used a mapping of 7 intrinsic connectivity networks derived by Yeo et al. (2011) to the high-resolution MMP atlas, to contextualise our results. This mapping, previously described and used in Váša et al. (2020), was obtained as follows. We first computed surface overlap (at the vertex level, using CIVET software; Ad-Dab'bagh et al., 2006) between each MMP atlas parcel and each intrinsic connectivity network, before assigning each MMP atlas parcel to the network that it overlapped most. Parcels of the high-resolution MMP atlas excluded from analyses due to limited FoV in EPImix scans (Fig. S3) were also excluded from intrinsic connectivity network visualisations and analyses.

Correspondence between EPImix and single-contrast T 1 -weighted scan intensities
Local correlations of T 1 -w intensities were generally positive. At the voxel level, correlations were highest in the grey matter and cerebrospinal fluid (Spearman's ρ ≤ 0.80), and lower in white matter (Fig. S5A,B). Within regions of interest of the MMP atlases, correlations were lower but predominantly positive, both at the high resolution (ρ ≤ 0.41; Fig. S5C) and at the low resolution (ρ ≤ 0.30; Fig. S5D).
We next quantified the within-and between-participant correspondence of EPImix and single-contrast data (Fig. S6A). We calculated global identifiability, as the difference of the median between-participant correlation and median withinparticipant correlation ( Fig. S6B; relevant parts of the correlation matrices are depicted in Fig. S6C). Identifiability was low at the level of brain voxels (I dif f = 0.62 -0.50 = 0.12), but considerably higher when correlating cortical GM voxels only (I dif f = 0.47 -0.23 = 0.24). Averaging intensities within regions of interest led to increases in both withinparticipant and between-participant correlations, resulting in decreased identifiability -both for the high-resolution atlas (I dif f = 0.61 -0.43 = 0.19) and the low-resolution atlas (I dif f = 0.78 -0.66 = 0.12). For regional data, we additionally used a null model relying on spherical "spin" permutation of cortical regions to account for spatial autocorrelation of the data when quantifying spatial correspondence between contrasts. Within the high-resolution atlas, 66/66 = 100% of within-participant correlations survived the FDR-corrected permutation test, compared to 3243/4290 = 75.6% of between-participant correlations. Within the low-resolution atlas, 64/66 = 97.0% of within-participant correlations survived the permutation test, compared to 3029/4290 = 70.6% of between-participant correlations ( Fig. S6A). Finally, we calculated individual-level identifiability, as the fraction of times that within-participant scan correlations are higher than between-participant scan correlations, using one of the contrasts as a reference (Fig. S6D). Individual identifiability was highly similar when using EPImix T 1 -w scans and T 1 -w scans as reference. Individual participants were most identifiable at the level of GM voxels, with high individual identifiability at the level of all brain voxels and regions of the high-resolution atlas as well; regions of the low-resolution atlas led to comparatively lower individual identifiability (Fig. S6D).
To dissect the effect of voxel-wise smoothing on acrossand between-participant correspondence as well as identifiability, we repeated a subset of the above analyses after smoothing voxel-wise data using 2, 4, and 6 mm FWHM kernels, and compared results to unsmoothed data (0 mm FWHM below) (Fig. S7). We first inspected the correlation, across participants, of all brain voxels as a function of smoothing kernel size. The effect of smoothing was to reduce correlations; median correlations decreased as a function of smoothing, both within the whole-brain mask (Md(ρ) for: 0 / 2 / 4 / 6 mm FWHM = 0.17 / 0.17 / 0.15 / 0.13), and within the GM mask (Md(ρ) for: 0 / 2 / 4 / 6 mm FWHM = 0.22 / 0.22 / 0.20 / 0.17) (Fig. S7A). We next investigated between-participant correspondence using voxel-wise GM T 1 -w intensities, which is the voxel-wise type of data for which identifiability was highest in unsmoothed data (I dif f = 0.24, compared to I dif f = 0.12 for all brain voxels). The effect of smoothing was to increase both within-participant and between-participant correlations, but with a greater increase in the latter; resulting in reduced differential identifiability as a function of increasing smoothing kernel size (I dif f for: 0 / 2 / 4 / 6 mm FWHM = 0.24 / 0.24 / 0.21 / 0.16; Fig. S7B,C).
Finally, we constructed networks of T 1 -w intensity covariance, using both EPImix and single-contrast scans. While single-contrast T 1 -w structural covariance networks showed similar hallmarks of organisation to structural covariance networks commonly constructed from regional cortical thickness or grey matter volume data, such as strong long-range inter-hemispheric correlations between homotopic regions, structural covariance networks constructed from EPImix data instead showed high short-range correlations, clustered in frontal cortex; particularly so for regions of the high-resolution atlas (Fig. S8). The correspondence between the upper triangular parts of the structural covariance matrices was modest for the high-resolution atlas (Spearman's ρ = 0.22), with higher correspondence for the low-resolution atlas (Spearman's ρ = 0.45).
processing time (s) quality (Dice) Step Evaluation  Table S1: Statistical details of the impact of processing steps on time and quality. For each pair of pipelines under comparison, we list the median within-participant difference in processing time, the median within-participant difference in quality (evaluated using the Dice coefficient), as well as the corresponding Wilcoxon signed-rank test raw P-value and Bonferroni-corrected P-value. Median differences in processing time and quality were calculated by subtracting values corresponding to the first pipeline from the second (within participants); i.e. for Evaluation "A | B", ∆ M d = Md(B − A). Figure S1: Age distribution of participants by sex and scan sequence. Scans from a total of 95 participants (48 female, 47 male) were included in this study. Of those, 66 (33 female, 33 male) were scanned using both EPImix and single-contrast T 1 -weighted sequences, while an additional 29 (15 female, 14 male) were scanned using EPImix only. There were no significant differences in participant age by sex or scan sequence (Chi-squared test, χ 2 = 0.005, P = 0.95).   Figure S4: Evaluation of regional quality of registration (and preceding steps) using the Mindboggle dataset. Left: Regional Dice coefficient values quantifying the overlap between "manually" registered atlas labels and those released with the Mindboggle dataset (Klein and Tourville, 2012), for each of seven evaluated processing pipelines (rows 1-3: spatial resolution; row 4: bias field correction; row 5: brain extraction; row 6: b-spline SyN registration; row 7: "reference pipeline" with a slower but higher quality version of the ANTs SyN registration algorithm). Right: Regional differences in Dice coefficient values. Pairs of maps being compared are joined by grey braces. Differences were calculated by subtracting values of the map below from the map above (i.e. ∆Dice = Dice above -Dice below ). Figure S5: Local correspondence of T 1 -w intensities across participants. Spearman's correlations between intensities of rapidlyprocessed T 1 -w scans from the EPImix sequence and a single-contrast acquisition, using data of 66 participants. Correlations are depicted: at the voxel level for A) the whole brain, and B) cortical grey matter, as well as within regions of interest of C) the high-resolution and D) the low-resolution MMP atlas. E) Distributions of correlations at each spatial resolution considered (as depicted in panels A-D). Note that few of these correlations were significant (respectively 20 / 31% of brain / GM voxels, and 3 / 0% regions of the high / low-resolution MMP atlases.) (At the regional level, median regional values were extracted prior to calculation of correlations for each region.) Figure S6: Participant identifiability across EPImix and single-contrast scans, using T 1 -w scan intensities. Between-participant correlations and identifiability were investigated using four types of input data, at three spatial resolutions (columns in panels A-B, rows in panel D): all brain voxels, cortical grey matter (GM) voxels, regions of the high-resolution MMP atlas, and regions of the low-resolution MMP atlas. A) Spearman's correlations between EPImix and single-contrast T 1 -w scan intensities, within and between participants. Cross-contrast correlations at the level of regions of interest were benchmarked using a null model controlling for contiguity and spatial autocorrelation (upper triangular blocks). B) Differential identifiability of contrasts, defined as the difference between the median within-participant correlation (right / red y-axes) and the median between-participant correlation (left / grey y-axes), as illustrated in C). D) Individual identifiability, defined as the fraction of times that the within-participant correlation is higher than between-participant correlations, either identifying a single-contrast T 1 -w scan relative to EPImix T 1 -w scans (EPImix T 1 -w ref.), or vice-versa (T 1 -w ref. ). P-values correspond to the (paired) Wilcoxon signed-rank test between neighbouring distributions (testing whether different spatial resolutions of data lead to differences in individual identifiability (H 1 ), or whether there is no statistical difference between these values (H 0 )). Figure S7: Effects of data smoothing on between-participant correspondence and identifiability of voxel-wise GM T 1 -w intensities. A) Spearman's correlations between voxel-wise T 1 -w intensities of rapidly-processed scans from the EPImix sequence and a single-contrast acquisition across 66 participants, as a function of smoothing. B) Spearman's correlations between EPImix and single-contrast T 1 -w scan intensities, within and between participants. C) Differential identifiability of contrasts, defined as the difference between the median within-participant correlation (right / red y-axes) and the median between-participant correlation (left / grey y-axes) (as illustrated in main text Fig. 5C and Fig S6C). Figure S8: Structural covariance networks constructed from EPImix and single-contrast T 1 -w intensities. A) Structural covariance networks constructed using the high-resolution MMP atlas (297 regions). The diamond plot (top) is ordered according to regional membership of the 7 canonical intrinsic connectivity networks derived by Yeo et al., 2011. Network diagrams depict the strongest 0.3% correlations. B) Structural covariance networks constructed using the low-resolution MMP atlas (32 regions). Network diagrams depict the strongest 10% correlations. Figure S9: Morphometric similarity networks constructed from EPImix contrasts and FreeSurfer reconstructions of singlecontrast T 1 -w scans. A) Group-average morphometric similarity networks constructed using the high-resolution MMP atlas (297 regions). The diamond plot (top) is ordered according to regional membership of the 7 canonical intrinsic connectivity networks derived by Yeo et al., 2011. Network diagrams depict the strongest 0.3% correlations. B) Group average morphometric similarity networks constructed using the low-resolution MMP atlas (32 regions). Network diagrams depict the strongest 10% correlations. Panels C) and D) depict the distribution of Spearman's ρ within individual participants. Figure S10: Correspondence between structural covariance and morphometric similarity networks derived from EPImix and single-contrast T 1 -w scans across intrinsic connectivity networks. Spearman's ρ between groups of edges within and between intrinsic connectivity networks derived by Yeo et al., 2011, in networks constructed using the high-resolution MMP atlas (297 regions). A) Correlations between structural covariance networks constructed from EPImix and single-contrast T 1 -w scan log-Jacobians (left), and T 1 -w intensities (right). B) Correlations between morphometric similarity networks constructed from EPImix contrasts and FreeSurfer reconstructions of single-contrast T 1 -w scans.