Leveraging the crowd for annotation of retinal images

Medical data presents a number of challenges. It tends to be unstructured, noisy and protected. To train algorithms to understand medical images, doctors can label the condition associated with a particular image, but obtaining enough labels can be difficult. We propose an annotation approach which starts with a small pool of expertly annotated images and uses their expertise to rate the performance of crowd-sourced annotations. In this paper we demonstrate how to apply our approach for annotation of large-scale datasets of retinal images. We introduce a novel data validation procedure which is designed to cope with noisy ground-truth data and with non-consistent input from both experts and crowd-workers.