A Two-Stage Framework with Ordinal Constraints for Age Estimation from Retinal Fundus Images

Predicting age from medical images is an important tool for assessing biological aging. However, conventional imaging methods such as MRI are costly and less accessible. To address this, we propose a deep learning framework based on retinal fundus images, which are more affordable and widely available. Our two-stage method introduces three key components: a pre-processing pipeline customized for fundus-specific characteristics, an ordinal ranking loss to enhance prediction consistency, and an early fusion strategy to integrate demographic data. Evaluated on a curated dataset of 11,434 high-quality fundus images, our model achieves a mean absolute error (MAE) of 4.202 years, outperforming current approaches and demonstrating strong potential for large-scale population aging assessment.