PRETI: Patient-Aware Retinal Foundation Model via Metadata-Guided Representation Learning

TL;DR

The paper tackles the limited availability of labeled retinal data by leveraging readily available metadata (age and gender) to improve retinal image understanding. It introduces PRETI, a metadata-guided retinal foundation model built on SiamMAE that fuses Learnable Metadata Embedding (LME) with Retina-Aware Adaptive Masking (RAAM) and patient-level data pairs to learn both global structures and fine-grained pathologies. The method optimizes a joint objective combining reconstruction, consistency, and meta losses, enabling dynamic metadata adaptation and robust cross-view representations. Empirical results on in-house and public datasets show state-of-the-art performance in disease and biomarker prediction, highlighting the practical impact of metadata-guided foundation models in retinal disease analysis; code and pretrained models are publicly available.

Abstract

Retinal foundation models have significantly advanced retinal image analysis by leveraging self-supervised learning to reduce dependence on labeled data while achieving strong generalization. Many recent approaches enhance retinal image understanding using report supervision, but obtaining clinical reports is often costly and challenging. In contrast, metadata (e.g., age, gender) is widely available and serves as a valuable resource for analyzing disease progression. To effectively incorporate patient-specific information, we propose PRETI, a retinal foundation model that integrates metadata-aware learning with robust self-supervised representation learning. We introduce Learnable Metadata Embedding (LME), which dynamically refines metadata representations. Additionally, we construct patient-level data pairs, associating images from the same individual to improve robustness against non-clinical variations. To further optimize retinal image representation, we propose Retina-Aware Adaptive Masking (RAAM), a strategy that selectively applies masking within the retinal region and dynamically adjusts the masking ratio during training. PRETI captures both global structures and fine-grained pathological details, resulting in superior diagnostic performance. Extensive experiments demonstrate that PRETI achieves state-of-the-art results across diverse diseases and biomarker predictions using in-house and public data, indicating the importance of metadata-guided foundation models in retinal disease analysis. Our code and pretrained model are available at https://github.com/MICV-yonsei/PRETI

Figures (2)

• Figure 1: The architecture of PRETI, a foundation model for retinal image representation learning, illustrates the integration of color fundus photography (CFP) and metadata through Learnable Metadata Embedding (LME). The model processes paired retinal images using Patient-level Pairs and employs shared encoders with dynamic metadata adaptation (age and gender). It also incorporates the Retina-Aware Adaptive Masking (RAAM) strategy for selective masking and reconstructing images via a decoder to enhance representation learning for robust clinical generalization.
• Figure 2: Visualization of attention maps for the [CLS] token and metadata embeddings, comparing focus with and without LME. These maps highlight regions key to diagnosing eye diseases like diabetic retinopathy and glaucoma.