Abstract:
Brain age prediction from biomechanical brain imaging is a promising approach for identifying
early signs of neurodegeneration. By estimating the deviation between predicted and chronological age, it is
possible to quantify accelerated brain aging, which can serve as a sensitive biomarker for disease risk. Biome
chanical properties derived from Magnetic Resonance Elastography (MRE) offer greater age sensitivity than
anatomical measures and can reveal early neurodegenerative changes in multicentre datasets. Consistent
performance across heterogeneous datasets poses a challenge, as demographic variability and acquisition-site
differences can introduce confounding effects that limit generalisation. To address this, we investigate con
trastive learning approaches tailored for regression, enabling models to capture fine-grained biomechanical
patterns associated with ageing. Furthermore, these are combined with and compared to different domain
adaptation strategies, including an adversarial classifier and distribution measure penalties such as Maximum
Mean Discrepancy (MMD) and Hilbert Schmidt independence criterion (HSIC), designed to reduce site- and
demographic-specific biases while preserving biologically meaningful variation. Our results show that the
model combining contrastive regression with MMD achieved the most accurate brain age estimation. How
ever, improvements in predictive accuracy did not consistently correspond to reductions in domain-specific
confounding. This underscores the importance of evaluating both accuracy and robustness when developing
generalisable biomarkers for brain ageing. Through this integration and systematic comparison, the work
aims to identify representation learning strategies that improve generalisation in brain age prediction, sup
porting the development of biomechanical imaging as a reliable biomarker in multi-centre studies.
