Brain age prediction from MRI scans using deep learning: A robust multi-site framework for neurodegenerative disease biomarker development

Brain age prediction has emerged as a promising computational biomarker for understanding healthy aging and identifying deviations associated with neurodegenerative diseases. This study constructed a deep learning-based brain age prediction model using magnetic resonance imaging scans, aiming to provide a reference for developing diagnostic biomarkers for Alzheimer's disease and Parkinson's disease. We developed and rigorously evaluated novel three-dimensional convolutional neural network models for predicting brain age from T1-weighted magnetic resonance imaging scans. The models were trained and validated on four heterogeneous open neuroimaging repositories (OASIS, IXI, ABIDE, and ABIDE II), providing a diverse multi-site corpus of 2 , 072 subjects aged 6 − 96 years. We adopted a minimal preprocessing strategy (resampling, cropping/padding, and mean-image subtraction) and introduced targeted augmentations, notably rotation and residual non-brain tissue inclusion, to systematically test robustness under realistic clinical perturbations. This model achieved competitive performance with r = 0.90 , and a root mean squared error of 3.66 years on held-out test data with augmentation, representing a substantial improvement over the 9.66 years root mean squared error observed without proper augmentation strategies and demonstrating the critical importance of robust training methodologies. The training root mean squared error reached 2.83 years, indicating excellent model fitting without overfitting. We present comprehensive analyses of robustness, interpretability, and clinical relevance. Elevated brain-age gaps produced by the model show anatomical plausibility for potential applications to neurodegenerative diseases, including Alzheimer’s disease and Parkinson’s disease. The study provides detailed methodological notes to support reproducibility and future validation in disease-labeled cohorts while contextualizing these results within the broader landscape of brain age prediction research and neurodegenerative disease biomarkers.