RETBind: a patient-level multimodal foundation model for unified cross-modal retinal representation learning

Medical artificial intelligence (AI) holds considerable promise across a wide range of ophthalmic tasks, offering a practical path to patient-level diagnostic support. Foundation models further extend this potential by learning generalizable representations from large-scale unlabeled data, alleviating the burden of extensive manual annotation. However, existing ophthalmic foundation models largely rely on single-modality or modality-specific encoders, failing to capture patient-level cross-modal correspondence in real-world multimodal clinical workflows. To address these gaps, we propose RETBind, a patient-level multimodal foundation model with a unified encoder that explicitly learns cross-modal correspondence from paired clinical imaging using patient-level paired optical coherence tomography (OCT) and ultra-widefield (UWF) fundus imaging. This patient-level paired pretraining strategy enables the model to learn unified cross-modal representations of complementary information, improv-ing performance and robustness beyond single-modality approaches. Across nine benchmark datasets spanning ocular disease classification and segmentation tasks, RETBind achieved consistent improvements, particularly enhancing AUPR in clinically relevant imbalanced settings. Exploratory analyses further examined the potential of RETBind for systemic disease prediction from retinal imaging, collectively highlighting its ability to learn unified cross-modal representations within a single shared encoder, enabling direct interaction between modalities and supporting transfer across diverse retinal imaging tasks.