Unsupervised Deep Learning Identifies a Latent Peripapillary Vascular Biomarker for MS in Fundus Photography

Early diagnosis of multiple sclerosis (MS) remains a challenge, delayed by the expensive imaging and invasive procedures. In our quest for a biomarker that aids in the diagnosis of MS, we investigated a novel diagnostic pathway for MS using an unsupervised deep-learning framework (RETFound) applied to routine 2D fundus photography. By comparing a physician-curated dataset of confirmed cases of MS as clinical ground truth against controls, our model autonomously identified the peripapillary microvasculature as a key diagnostic signature. The RETFound framework assigned 76.8% of the total decision-making weight to a specific anatomical substrate representing the en face integration of four-tiered capillary plexuses. The proficiency of this 'vascular fingerprint' was established through an independent hold-out validation partition (n = 244, 122 MS, 122 controls), achieving an AUC of 0.97 and demonstrating that 2D retinal pixels harbor a latent signature of MS that is detectable even in the absence of explicit anatomical guidance. By extracting a high-fidelity signal from a ubiquitous and low-cost modality, we establish a scalable screening and diagnostic framework. Our results suggest that the integration of microarchitectural vascular data into routine clinical workflows in the appropriate clinical context could potentially transform MS diagnosis, after exclusion of mimics. To our knowledge, this study represents the inaugural use of unsupervised deep learning on a curated MS cohort, identifying the peripapillary vascular network as a distinctive pathobiological signature of the disease. These findings could be foundational in identifying clinically relevant vascular biomarkers in MS.