Dijsselhof, Mathijs BJ;
Moore, Candace;
Amiri, Saba;
Tee, Mervin;
Hilal, Saima;
Chen, Christopher;
van den Born, Bert-Jan H;
... Petr, Jan; + view all
(2025)
Multi-cohort, multi-sequence harmonisation for cerebrovascular brain age.
Imaging Neurosci (Camb)
, 3
, Article IMAG.a.964. 10.1162/IMAG.a.964.
Preview |
Text
Hughes_imag.a.964.pdf Download (1MB) | Preview |
Abstract
Higher brain-predicted age gaps (BAG), based on anatomical brain scans, have been associated with cognitive decline amongst elderly participants. Adding a cerebrovascular component, in the form of arterial spin labelling (ASL) perfusion MRI, can improve the BAG predictions and potentially increase sensitivity to cardiovascular health, a contributor to brain ageing and cognitive decline. ASL acquisition differences are likely to influence brain age estimations, and data harmonisation becomes indispensable for multi-cohort brain age studies including ASL. In this multi-cohort, multi-sequence study, we investigate harmonisation methods to improve the generalisability of cerebrovascular brain age. A multi-study dataset of 2608 participants was used, comprising structural T1-weighted (T1w), FLAIR, and ASL 3T MRI data. The single scanner training dataset consisted of 806 healthy participants, age 50 ± 17, 18-95 years. The testing datasets comprised four cohorts (n = 1802, age 67 ± 8, 37-90 years). Image features included grey and white matter (GM/WM) volumes (T1w), WM hyperintensity volumes and counts (FLAIR), and ASL cerebral blood flow (CBF) and its spatial coefficient of variation (sCoV). Feature harmonisation was performed using NeuroComBat, CovBat, NeuroHarmonize, OPNested ComBat, AutoComBat, and RELIEF. ASL-only and T1w+FLAIR+ASL brain age models were trained using ExtraTrees. Model performance was assessed through the mean absolute error (MAE) and mean BAG. ASL feature differences between cohorts decreased after harmonisation for all methods (p < 0.05), mostly for RELIEF. Negative associations between age and GM CBF (b = -0.37, R2 = 0.13, unharmonised) increased after harmonisation for all methods (b < -0.42, R2 > 0.12), but weakened for RELIEF (b = -0.28, R2 = 0.14), In the ASL-only model, MAE improved for all harmonisation methods from 11.1 ± 7.5 years to less than 8.8 ± 6.2 years (p < 0.001), while BAGs changed from 0.6 ± 13.4 years to less than -1.03 ± 7.92 years (p < 0.001). For T1w+FLAIR+ASL, MAE (5.9 ± 4.6 years, unharmonised) increased for all harmonisation methods non-significantly to above 6.0 ± 4.9 years (p > 0.42) and significantly for RELIEF (6.4 ± 5.2 years, p = 0.02), while BAGs non-significantly differed from -1.6 ± 7.3 years to between -1.3 ± 4.7 and -2.0 ± 8.0 years (p > 0.82). In general, the ASL-specific parameter harmonisation method AutoComBat performed nominally best. Harmonisation of ASL features improves feature consistency between studies and also improves brain age estimations when only ASL features are used. ASL-specific parameter harmonisation methods perform nominally better than basic mean and scale adjustment or latent-factor approaches, suggesting that ASL acquisition parameters should be considered when harmonising ASL data. Although multi-modal brain age estimations were improved less by ASL-only harmonisation, possibly due to weaker associations between age and ASL features compared with T1w features importance, studies investigating pathological ASL-feature distributions might still benefit from harmonisation. These findings advocate for ASL-parameter specific harmonisation to explore associations between cardiovascular risk factors, brain ageing, and cognitive decline using multi-cohort ASL and cerebrovascular brain age studies.
Archive Staff Only
 |
View Item |
