The value of arterial spin labelling perfusion MRI in brain age prediction

Background: Biological brain age estimates from structural MRI data and their difference from chronological age — the brain age gap (BAG) — have been successfully applied in a range of diseases, but remain limited to capturing structural characteristics only. To incorporate physiological properties, we operationalized ‘Cerebrovascular brain age’ using a combination of structural, and arterial spin labelling (ASL) image data, investigate their optimal feature and algorithm combinations, and evaluate its repeatability.
Methods: Healthy participants (n = 341, 62 % female, age 59.7 ± 14.8 years, range: 21 - 95 years) were scanned at baseline and after 1.7 ± 0.5 years (n = 248, 62.9 % female, mean age 62.4 ± 13.3 years, range: 27 - 86). At 3 T MRI, 3D structural T1-weighted (T1w) and Fluid Attenuated Inversion Recovery (FLAIR), and 3D ASL image data were acquired to extract within grey matter (GM) and deep white matter (WM) ROIs: volumetrics, WM hyperintensity volume and count; and cerebral blood flow (CBF) and spatial coefficient of variation (CoV). Multiple combinations of features and machine learning algorithms were evaluated to train brain age algorithms on 70 % of the subjects and evaluated on the remainder, for 300 Monte-Carlo cross-validations, using the Mean Absolute Error (MAE). Feature importance of the best performing model was assessed by determining the feature weights. Model repeatability of the best model was assessed by comparing the BAGs between baseline and follow-up, also using T1w + FLAIR or ASL-only features.
Results: The lowest MAE was observed for the ElasticNetCV algorithm using T1w + FLAIR + ASL (MAE = 5.03 ± 0.34 years) and significantly better compared to using T1w + FLAIR (MAE = 6.01 ± 0.39, p < 0.01) and ASL-only features (MAE = 6.04 ± 0.39, R2 = 0.70 ± 0.04, p < 0.01). The three most important features were GM CBF (6.2 ± 1.18), GM/ICV (5.34 ± 0.6), and WM CBF (4.16 ± 0.36).
Average baseline and follow-up BAGs were not different (-1.51 ± 6.29 and -1.14 ± 6.40 years respetively, ICC = 0.85, 95% CI: 0.79 - 0.90, p = 0.16). The ElasticNetCV model with T1w+FLAIR+ASL performed similar to the same model with the T1w + FLAIR feature set (0.37 ± 3.48 years and 0.01 ± 2.95 years respectively, p = 0.14), and the ASL-only feature set (0.29 ± 4.03, p = 0.39).
Conclusion: The addition of ASL features to structural brain age improved brain age prediction, with the ElasticNetCV algorithm and a combination of all tested features (T1w+FLAIR+ASL) performing the best in a cross-sectional and repeatability comparison. These findings encourage future studies to explore the value of ASL in brain age in various pathologies.