Feature Space Translation Framework for Cross-Device Alignment and Derivation of Clinically Relevant Retinal Biomarkers from Foundation Model Embeddings

Sparsh Rastogi, Akshat Bakshi, Jidugu Sriharisesh, Jatin Bedi

Keywords: Oculomics, Foundation Models, Latent Space Alignment, Retinal Imaging, Machine Learnin

Abstract:

Oculomics enables non-invasive assessment of systemic diseases through retinal imaging, offering a portable, cost-effective alternative to traditional methods. However, its real-world deployment is limited by device-induced variability and poor generalizability across heterogeneous imaging systems. In recent years, large retinal foundation models such as RetFound and RetFound Green have shown strong performance on several downstream tasks. However, their generalization across imaging devices remains limited. Furthermore, the high-dimensional representations produced by these models do not directly correspond to clinically meaningful retinal biomarkers, further limiting their interpretability and clinical utility.

To address these challenges, we propose a feature space translation framework using tabular neural networks to align feature distributions across devices, thereby improving cross-device generalization. We further investigate whether the same framework can be used to translate final-stage foundation model embeddings into clinically meaningful retinal biomarkers, enabling more interpretable representations.
Experiments on the AI-READI dataset show that our framework supports cross-device translation of retinal biomarkers, enables biomarker extraction from foundation model embeddings, and allows cross-device biomarker estimation directly from embeddings. Together, these results demonstrate seamless integration of our approach into existing oculomics pipelines while improving the consistency, interpretability, and clinical usability of retinal biomarkers.

Submission Number: 167