Cerebrovascular Brainage

Target audience
Physicists aiming to use Brain-age in ASL. Clinicians looking at healthy and accelerated aging.

Purpose
Brain-age estimates the biological brain age from structural MRI images based on changes in brain-tissue integrity and irreversible structural changes [1]. The brain-age offset to the chronological age — the age gap — is associated with cognitive pathology [2]. Adding vascular or functional MRI biomarkers may add sensitivity to physiological and metabolic changes, complementing and improving structural brain-age, and possibly improving its sensitivity to earlier disease changes. Arterial spin labeling (ASL) MRI is a potential early biomarker of cerebrovascular health and correlates with the initial stages of cognitive pathology [3]. Here, we propose the ‘Cerebrovascular brain-age’ as a combination of T1w, FLAIR, and ASL image features composed of the spatial Coefficient of Variation (CoV) and vascular-territory derived (VT) cerebral blood flow (CBF).

Methods
Baseline and 1.7±0.5y follow-up T1w, FLAIR, and 3D PCASL data of 233 healthy participants — 36.8% male, 61.4±13.5y — were drawn from the Norwegian Centre for Mental Disorders Research StrokeMRI dataset and processed with ExploreASL. We compared different features (Figure 1A), feature sets (Figure 1B), and machine learning algorithms readily available within Python packages (Figure 1C). As deep WM CBF was not associated with age, we used it to normalise all CBF features to correct for individual physiological differences. Model performance was assessed for each algorithm and feature set combination, using mean absolute error (MAE) of the age gaps and 300 Monte-Carlo cross-validation simulations using a 70:30 train-test split (n=326; n=140). The best-performing model was used in subsequent analyses. Model performance was compared to brainageR (trained on 3337 healthy individuals [4]) using a single 70:30 train-test split.
The best-performing model was then assessed for robustness by studying the age gap differences between baseline and follow-up in the same train-test split, compared using a paired t-test. Lastly, we compared the age gaps of the best-performing model and T1w feature set using the paired t-test.

Results
The ExtraTrees algorithm showed the lowest MAE for T1w+FLAIR+ASL (MAE = 4.47) (Figure 2). On a single split, this model (MAE=5.36) outperformed brainageR (MAE = 5.67), however not significantly (p = 0.06). Our model showed good reproducibility, the average baseline and follow-up age gaps were similar, respectively -1.18 ± 6.7 and -1.31 ± 6.2 years (p=0.38) with an average difference of 0.135 ± 3.8 years only (ICC = 0.83) (Figure 3). Reproducibility of T1w+FLAIR+ASL features was equally good as for only T1w features (p=0.17).

Discussion
The ExtraTrees model utilising ASL, FLAIR, and T1w features improved the prediction of brain-age the most and showed good long-term reproducibility. Significant improvement from the existing brainageR model was not reached; probably due to differences in sample size and age distribution. Furthermore, our model slightly overestimated young age and underestimated old age, which is the opposite of brainageR. Despite our relatively small population and the physiological variability of CBF, our cerebrovascular brain-age prediction remained consistent from baseline to follow-up. Our limitations include the relatively small, single-site, training and validation set. Future studies are encouraged to improve generalisation and to determine correlations between Cerebrovascular brain-age and cognitive decline.

Conclusion
The addition of ASL features to structural brain-age improved brain age prediction across algorithm types. The ExtraTrees algorithm performed best, however not significantly different from brainageR, and showed good longitudinal reproducibility.