MRI signatures of Brain Age in the Alzheimer’s disease continuum

Abstract

AbstractBackgroundAge is the biggest risk factor for Alzheimer’s disease (AD). Substantial efforts to extract age signatures from magnetic resonance imaging (MRI) of the brain have achieved impressive accuracy and demonstrated these signatures were altered by neurological disorders including AD. In this work, we develop a deep learning model to characterize brain age signatures and examine their variation over the AD continuum.MethodA 3D deep ResNet model was implemented to learn the structural signatures of brain age in predicting the chronological age of the participants. The model was trained on 7372 T1‐weighted MRI scans from a combined lifespan cohort of 5848 cognitively normal participants (age: 8‐ 95 yrs) using a 10‐fold cross validation procedure. The model was then applied to the independent ADNI cohort (N = 1175, 50‐98 yrs) to examine MRI derived brain age signatures variation in the AD continuum. The difference (∆age) between model predicted age and chronological age was used as the independent variable for group comparison using univariate ANCOVA tests, controlling for sex, education, and APOE4 gene dose, among five clinical groups: normal control (NC), normal to MCI converters (NC‐MCI), stable MCI (MCIs), MCI to AD converters (MCI‐AD), and AD patients.ResultThe ResNet model achieved an MAE = 3.76 yrs and R2 = 0.93 across 10 folds in the lifespan cohort. In the baseline data of the ADNI cohort, model derived ∆age increased along the AD continuum as expected: NC (‐1.2yrs) < NC‐MCI (‐0.7yrs) < MCIs (‐0.3yrs) < MCI‐AD (0.7yrs) <AD (1.5yrs) (F = 11.24, p<0.0001). Post‐hoc pairwise comparison results are summarized in Table 1/Figure 2. Although the difference between NC and NC‐MCI did not reach significance (p = 0.24) for the baseline data, comparison between the two groups using data from all visits was significant (p = 0.0009), suggesting more subtle differences.ConclusionA ResNet model was developed that captures the MRI signatures of brain age with high accuracy in age prediction. The model detected group level differences along the AD continuum in the expected direction. Subtle changes in brain signatures were observable in normal participants who later developed cognitive impairment. Our technique provides a novel approach to investigate the relationship between aging and AD.