Latent space structure of retinal age: an artificial intelligence–based analysis of aging in fundus images

Abstract

Abstract This study investigated whether age-related variation is organized as a geometric structure in latent space and how fine-tuning reshapes this representation. Fundus photographs from fellow eyes of patients who underwent macular surgery were retrospectively analyzed. After quality control, 4,602 images (885 eyes) were split into training and test sets by age. A retinal age prediction model trained on the Japan Ocular Imaging Registry was fine-tuned. Latent-space geometry was evaluated using principal component analysis and Uniform Manifold Approximation and Projection, and geometric metrics. Transformer and convolutional neural network models were compared, and spatial attention was evaluated using occlusion analysis. The pretrained model exhibited a ceiling effect with underestimation at older ages. Fine-tuning improved age alignment, increasing correlations and coefficient of determination while stabilizing local slope. Age-related variation formed a nonlinear continuous manifold rather than a linear axis, and the ceiling effect was consistent with reduced age-related sensitivity or structural convergence at higher ages. Across models, transformer models showed clearer low-dimensional structure, whereas convolutional neural network models showed more dispersed representations despite similar predictive performance. Fine-tuning also reduced peripheral dominance and increased anatomically relevant attention. These findings indicate that age-related representation in fundus images is geometrically organized and reshaped by fine-tuning.