Learning age-conditioned brain atlases via free-prototype modeling for brain age prediction.
Authors
Affiliations (2)
Affiliations (2)
- Institute of Computer Vision and Robotics, University of Girona, Girona, Catalonia, Spain. Electronic address: [email protected].
- Institute of Computer Vision and Robotics, University of Girona, Girona, Catalonia, Spain.
Abstract
Brain age (BA) estimation from structural MRI is a promising biomarker for the early detection, monitoring, and risk stratification of neurodegenerative disorders, yet most existing approaches rely on black-box deep learning models that limit clinical interpretability. Prototype-based learning offers a promising alternative by grounding predictions in reference representations, but current methods depend on selecting prototypes from real training samples, introducing subject-specific variability and requiring curated datasets. In this work, we propose a free-prototype framework that directly learns age-conditioned prototypes in latent space instead of anchoring them to individual training samples. These prototypes are decoded into age-conditioned brain atlases that provide clinically interpretable visual references of anatomical aging. Extensive experiments across multiple large-scale neuroimaging cohorts demonstrate that the proposed approach achieves BA predictive performance comparable to state-of-the-art convolutional models in-domain while improving robustness under distribution shift in external datasets. Moreover, our results show that deviations from age-matched prototypes capture clinically relevant information beyond the conventional BA gap, enabling improved discrimination between healthy controls and Alzheimer's disease patients. Hence, beyond the limited notion of accelerated aging as a biomarker of disease, our findings support the use of residual information as complementary evidence to improve diagnostic frameworks in clinical settings. Finally, structural analyses confirm that the learned atlases follow smoother and more biologically consistent aging trajectories than prototype representations based on real-image selection. These findings demonstrate that structuring the latent space around learned prototypes enables accurate and robust BA estimation while providing clinically interpretable normative references for understanding deviations from healthy brain aging.