Unsupervised discovery of latent retinal microarchitectural biomarkers in multiple sclerosis via gated attention multiple instance learning

The diagnosis of multiple sclerosis (MS) currently relies on complex workflows involving MRI and cerebrospinal fluid analysis, creating significant barriers to early detection. We present a weakly supervised deep learning framework utilizing unsupervised feature extraction capable of identifying previously occult retinal microarchitectural biomarkers of MS. By integrating a large-scale retinal foundation model (RETFound) with a gated attention multiple instance learning (MIL) architecture, we demonstrate that routinely acquired images contain a high-fidelity structural signal capable of distinguishing MS from controls with an AUC of 0.97 and accuracy of 91.4%. Critically, the model’s feature extraction was fully autonomous and requiring no a priori guidance regarding retinal anatomy. This indicates that the diagnostic signal emerges from intrinsic structural properties of the retina rather than human-imposed priors. Quantitative analysis of learned attention weights reveals that 76.8% of the decision-making mass is concentrated within localized patches in the peripapillary region, providing algorithmically-derived corroboration of MS-related neurodegeneration. This engineering approach bypasses the need for specialized optical coherence tomography (OCT), establishing a scalable, high-precision computational foundation for population-scale neurodegenerative screening.