What's next in AI for glaucoma screening? The REFUGE challenge outcomes

Abstract

Grand challenges provide a uniform training and evaluation framework to solve problems with AI-based developments. In ophthalmology, these challenges have allowed to boost research in fundus image analysis, particularly in diabetic retinopathy, where AI tools have achieved enough maturity to be even deployed in the clinics. Unfortunately, similar initiatives were never introduced for glaucoma classification, specially due to the difficulty of building large enough data sets with reliable annotations. To overcome this limitation we introduced REFUGE: the Retinal Fundus Glaucoma Challenge, held in conjunction with MICCAI 2018. The challenge consisted of two primary tasks, namely optic disc/cup segmentation and glaucoma classification. A data set of 1200 fundus images was publicly released to this end, including their ground truth segmentations and clinical glaucoma labels, as retrieved from clinical records. 12 teams qualified and participated in the online challenge, and their resulting algorithms were uniformly trained, evaluated and compared using our standardized evaluation platform. In this talk we will summarize the outcomes of the challenge and the lessons learned from its design, discussing also the expected characteristics of future challenges in the field.