Using deep learning transformer models to genotype/phenotype relationships in age-related macular degeneration (AMD) in NICOLA

Age-related macular degeneration (AMD) is a significant cause of vision loss, affecting millions worldwide. Wet AMD (wAMD), the advanced stage of AMD, can result in complete vision loss in both eyes if it develops in one eye. Early prediction of wAMD conversion in the unaffected eye is essential for timely clinical intervention. This study aims to predict wAMD conversion in the normal eye using deep learning, based on the superior temporal capabilities of Transformer models. We leveraged RetFound, a state-of-the-art Transformer model for retinal imaging, and fine-tuned it on OCT B-scans from a cohort of 3,330 patients. The model was trained to classify the absence or presence of specific phenotypes in OCT images and, in the case of presence, further distinguish between one or two copies of the phenotype. Preliminary experiments demonstrated that RetFound achieved approximately 70% accuracy in classifying OCT images, indicating its potential in identifying key genetic information. Our findings suggest that Transformer architectures, including RetFound, offer promise in predicting wAMD conversion based on specific phenotypes, providing a valuable tool for early detection and clinical decision-making.