Paper Title
OCT-to-Fundus Knowledge Distillation With Open Set Recognition for Retinal Disease Detection

Abstract
Diabetic retinopathy, glaucoma, and age-related macular degeneration are the most common examples of retinal disorders that are major causes of irreversible visual impairment in the world and demand early and reliable diagnosis methods. Despite providing better micro-structural information, the accessibility of Optical Coherence Tomography (OCT) is limited, and it cannot be deployed on a large scale in a largescale screening setting (where fundus photography is the norm). The recent studies have also demonstrated that cross-modal knowledge distillation has the capability of transferring the OCTderived depiction to models founded on the fundus to improve the level of diagnostic accuracy without necessitating OCT in the inference. In our current research, we build upon a known cross-modal distillation approach by using two clinical inspired changes. To begin with, we incorporate the Class-Balanced Focal Loss (CBFL) to the distillation pipeline in order to specifically replace the imbalance in classes and increase the sensitivity to the minority disease groups. Second, we also introduce an Open-Set Recognition (OSR) aspect that allows the system to be aware of the unknown retinal cases, thus reducing the overconfident error cases in the actual deployment. All these augmentations significantly enhance the strength, fairness and clinical dependability of fundus-only retinal screening systems. Keywords - Deep Learning, Fundus and OCT Images, Global Prototype Distillation (GPD), Knowledge Distillation, Local Concept Distillation (LCD), Open Set Recognition, Retinal Disease Classification, Vision-Language Model -FLAIR Model