Robust Optic Disc Localization by Large Scale Learning

Abstract

Since the optic disc (OD) is a main anatomical structure in retina, the localization of OD is an essential task in screening and diagnosing ophthalmic diseases. Many studies have been done for the automatic localization of OD but not reach a perfect performance yet. The bottleneck is lack of data and corresponding models that can handle with such big data. In this paper, we proposed an automatic OD localization method based on the hourglass network referenced from the human pose estimation task. Considering the lack of retina image databases, we also created a large retinal dataset of 85,605 images with manual OD bounding boxes. By learning from the large dataset, our deep network demonstrates excellent performance on OD localization. We also validated the proposed model on two public benchmarks, i.e. Messidor and ARIA datasets. Experiments show that it can achieve 100% accuracies on both datasets which clearly outperforms all the state-of-the-arts.