Repeat-consistency learning for improving repeatability of brain-age prediction from structural MRI.

Structural magnetic resonance imaging (MRI)-based brain-age prediction is increasingly used as an imaging-derived marker of brain ageing, but good predictive accuracy does not guarantee stable estimates when the same individual is scanned repeatedly. This study proposes repeat-consistency learning (RCL), a training framework that uses same-subject, same-session repeated MRI acquisitions as paired training information. A compact three-dimensional convolutional neural network was trained using a chronological-age prediction loss together with output-level and latent feature-level consistency losses. An unweighted formulation treated all repeated-scan pairs equally, whereas a disagreement-weighted formulation reduced the influence of pairs with larger predicted-age differences. The framework was evaluated using the Open Access Series of Imaging Studies cross-sectional dataset with subject-level training, validation and held-out test partitions. Both RCL formulations showed numerically lower scan-rescan disagreement and narrower Bland-Altman limits of agreement than prediction-only training, while held-out age-prediction performance remained broadly comparable. The unweighted formulation provided the most favourable overall balance between repeatability and predictive accuracy. Repeat-consistency learning also reduced prediction shifts under several controlled noise and affine perturbations, although substantial sensitivity to image blurring remained. These findings provide proof-of-concept evidence that repeated MRI acquisitions can be incorporated directly into model training to encourage more stable predictions and latent representations while maintaining broadly comparable chronological-age prediction performance. The observed improvements were modest and were demonstrated in a controlled, single-dataset, within-session setting. Validation in larger, multi-site, multi-scanner and longitudinal datasets is therefore required to determine the broader robustness, generalisability and practical utility of the approach.