Robust, fully-automated assessment of cerebral perivascular spaces and white matter lesions: a multicentre MRI longitudinal study of their evolution and association with risk of dementia and accelerated brain atrophy

Abstract

Background: Perivascular spaces (PVS) on brain MRI are surrogates for small parenchymal blood vessels and their perivascular compartment, and may relate to brain health. However, it is unknown whether PVS can predict dementia risk and brain atrophy trajectories in participants without dementia, as longitudinal studies on PVS are scarce and current methods for PVS assessment lack robustness and inter-scanner reproducibility.

Methods: We developed a robust algorithm to automatically assess PVS count and size on clinical MRI, and investigated 1) their relationship with dementia risk and brain atrophy in participants without dementia, 2) their longitudinal evolution, and 3) their potential use as a screening tool in simulated clinical trials. We analysed 46,478 clinical measurements of cognitive functioning and 20,845 brain MRI scans from 10,004 participants (71.1 ± 9.7 years-old, 56.6% women) from three publicly available observational studies on ageing and dementia (the Alzheimer's Disease Neuroimaging Initiative, the National Alzheimer's Coordinating Centre database, and the Open Access Series of Imaging Studies). Clinical and MRI data collected between 2004 and 2022 were analysed with consistent methods, controlling for confounding factors, and combined using mixed-effects models.

Findings: Our fully-automated method for PVS assessment showed excellent inter-scanner reproducibility (intraclass correlation coefficients >0.8). Fewer PVS and larger PVS diameter at baseline predicted higher dementia risk and accelerated brain atrophy. Longitudinal trajectories of PVS markers differed significantly in participants without dementia who converted to dementia compared with non-converters. In simulated placebo-controlled trials for treatments targeting cognitive decline, screening out participants at low risk of dementia based on our PVS markers enhanced the power of the trial independently of Alzheimer's disease biomarkers.

Interpretation: These robust cerebrovascular markers predict dementia risk and brain atrophy and may improve risk-stratification of patients, potentially reducing cost and increasing throughput of clinical trials to combat dementia.

Funding: US National Institutes of Health.

Keywords: Alzheimer’s disease; Dementia; Glymphatic system; Perivascular spaces; Small vessel disease; White matter lesions.

Fig. 1

Flowchart diagram depicting study design. Further information about the participants clinical and demographic information can be found in Table 2 and Supplementary Tables S5 and S7.

Fig. 2

Comparison between the original approach and our novel approach. The original approach requires to set a single threshold (black vertical dashed line in panel a, in this case equal to 10⁻⁴), and to use this threshold to segment vessel-like structures (Panel a). However, this approach lacks inter-scanner reproducibility, because the scale of the vesselness maps depends on the signal intensity of the input image, which may differ among MRI scanners and protocols. In fact, in this example the threshold would lead to very different number of segmented voxels: vessel-like masks from Siemens scanner will have approximately 12.5 and 25% more voxels than those from Philips and GE scanners, respectively (Panel b). In our novel approach, we set specific thresholds to the individual images (vertical dashed lines, red for the GE-derived image, green for the Philips-derived image, and blue for the Siemens-derived image) based on the value of the voxel corresponding to the 85th percentile (black horizontal solid line) of the total number of non-zero voxels of the vesselness map (Panel c). This approach leads to consistent PVS masks (Panel d) and robust metrics derived from the PVS masks (Supplementary Figs. S3a–c and S5), while preserving inter-individual differences and accuracy (Supplementary Figs. S3d–f and S4). Panels b and d report the 3D representations of the PVS segmentation masks of the same participant obtained with 3 different MRI scanners with fixed threshold approach (panel b) and our novel percentile-based approach (panel d). The small brain icon on the top left of each panel indicates the orientation of the PVS masks.

Fig. 3

Forest plots and spline plots for the associations of PVS and WML markers with dementia risk. In two-stage pooled analyses that combined individual-participant data from three studies (Panels a–f), each additional 100 WM-PVS (Panel a) and 10 BG-PVS (Panel b) were associated with 20% and 8% decrease in dementia risk, each additional 0.1 mm increase in mean WM-PVS (Panel c) and BG-PVS (Panel d) diameter were associated with 8% and 15% increase in dementia risk, and each unit increase of the log-transformed P-WML volume (Panel e) was associated with 13% increase in dementia risk. Log-transformed D-WML volume (Panel f) was not associated with dementia risk. In each graph, the size of the squares indicates the weight given to the study, and the width of the diamond indicates the 95% confidence interval for the overall association estimate. Between-study heterogeneity was statistically assessed with the use of I². The spline analysis of pooled data (Panels g–l) supported a linear association over the range of WM-PVS count (Panel g; 2.5th^–97.5th percentile, 206–762), BG-PVS count (Panel h; 2.5th–97.5th percentile, 71–244), WM-PVS diameter (Panel i; 2.5th–97.5th percentile, 1.77–2.30 mm), BG-PVS diameter (Panel j; 2.5th–97.5th percentile, 1.44–1.79) and P-WMH volume (Panel k; 2.5th–97.5th percentile, 0–9.1) within the overall population. Shaded areas indicate 95% confidence intervals, and the red line at 1.0 indicates the reference. Box plots at the bottom of the graphs show the distributions of the marker. The vertical bar indicates the median, and the ends of the box the interquartile range; the whiskers extend to values no farther than 1.5 times the interquartile range (which may be past the graphed area). P indicate P-values from the Cox models (panels a–f) and from the chi-square test for linearity (panels g–l). See also Supplementary Fig. S6 and Table S6 for sensitivity analysis. Data from three studies — the Alzheimer’s Disease Neuroimaging Initiative (ADNI), the Open Access Series of Imaging Studies (OASIS), and the National Alzheimer’s Coordinating Centre (NACC) — are shown. Hazard ratios were estimated from Cox models stratified according to study cohort with adjustment for age, sex, race, educational level, body mass index, CDR global score at the baseline, history of diabetes, cardio-/cerebro-vascular disease, hypertension, dyslipidaemia, family history of dementia, intracranial volume and the time interval between the MRI scan and the clinical visit of the cognitive assessment at the baseline.

Fig. 4

Plots for the estimated trajectories of brain atrophy in relation to PVS and WML markers. Top rows in each panel depict the effect of baseline PVS and WML markers on the trajectories of grey matter volume, cortical thickness, and white matter volume (Panels a–c, respectively). For each marker, equally spaced values from the low-risk (red), medium-risk (blue), and high-risk (yellow) tertile are shown (tertile limits in Supplementary Table S16). Shaded areas indicate 95% confidence intervals. P indicates adjusted significance value of the interaction term “marker” by “time” in linear mixed-effects models. The estimated volume or thickness preserved/lost per year for each additional unit increase in the marker are reported in Supplementary Table S8. The regional analysis (bottom rows in each panel) across cortical parcellations according to the Desikan-Killiany atlas reports the estimated volume or thickness preserved (positive values in red) or lost (negative values in blue) per year for each additional unit increase in the vascular marker. Only estimated values for regions that remained statistically significant after correction for multiple comparisons (68 comparisons) are shown; non-significant regions are greyed out. Coefficients for individual regions are reported in Supplementary Tables S9–S11 for grey matter volume, cortical thickness, and white matter volume, respectively. Estimates and corrected significance obtained from multivariable linear mixed-effects models with random intercepts and slopes for each individual participant (N = 3389 and 14,229 timepoints MRI scans). All models were adjusted for age, sex, race, educational level, body mass index, CDR global score at the baseline, history of diabetes, cardio-/cerebro-vascular disease, hypertension, dyslipidaemia, family history of dementia, intracranial volume, the baseline value of the dependent variable (grey matter volume, cortical thickness, or white matter volume), field strength, manufacturer, and intra-individual consistency of the protocol used for the longitudinal MRI acquisitions (consistent versus non-consistent protocol). See also Supplementary Figs. S10–S12 and Table S12 for sensitivity analysis.

Fig. 5

Plots for the estimated longitudinal trajectories of PVS or WML markers according to conversion to dementia status. Estimated longitudinal trajectories of WM-PVS (a–b), BG-PVS (c–d), and WML (e) markers for participants without dementia who converted to dementia (converters, red) and those that did not convert to dementia (non-converters, cyan). Trajectories are estimated from fully adjusted linear mixed-effect models; the adjusted P-values (P) indicate whether the trajectories are significantly different between converters and non-converters (interaction term “cognitive status” by “time” in linear mixed-effects models) after correction for multiple comparisons. Shaded areas indicate 95% confidence intervals. For WM-PVS markers (panels a–b), we estimated in each lobe of the left (L) and right (R) hemispheres the group-effect (expressed as T-value) for the longitudinal trajectories of the corresponding marker: positive values in red indicate significantly higher (i.e., less negative) slopes for non-converters versus converters, whereas negative values in blue indicate significantly lower (i.e., more negative) slopes for non-converters versus converters. Lobes where the longitudinal trajectories were not significantly different between converters and non-converters after correction for multiple comparisons are greyed-out. Estimates and corrected significance obtained from multivariable linear mixed-effects models with random intercepts and slopes for each individual participant (N = 3389 and 14,229 timepoints MRI scans). See also Supplementary Fig. S17 for sensitivity analysis.

Fig. 6

Error-bar plot for the relative sample size in simulated clinical trials enriched using PVS or WML markers. Relative sample sizes for simulated clinical trials in participants without dementia pooled from three studies (the Alzheimer’s Disease Neuroimaging Initiative, the Open Access Series of Imaging Studies, and the National Alzheimer’s Coordinating Centre). The simulations had statistical power of 80% at α = 0.05 and assumed a 30% treatment effect for slopes in cognitive decline, 1:1 allocation of treatment, total trial length of 48 months, and outcome measures every 12 months. All available longitudinal cognitive data (40,307 cognitive assessments from 7518 participants without dementia) were used in these multivariable models. Mean relative sample sizes and the corresponding standard errors of the mean are across 500 bootstrap iterations. The reference model (without enrichment and with 100% inclusion) included all the tertiles. In the two enrichment models for each marker (“medium-risk + high-risk tertiles”, red bars; “high-risk only tertile”, blue bars), only participants in the indicated tertiles were included. Tertile limits for each marker are reported in Supplementary Table S16. See also Supplementary Figs. S18 and S19 for additional simulations.