Multiple graph reasoning network for joint optic disc and cup segmentation

Glaucoma is one of the most irreversible eye diseases that causes visual damage worldwide. For glaucoma screening, it is essential to accurately segment the optic disc (OD) from the optic cup (OC) in fundus images. However, most of the known segmentation methods are two-stage processes and cannot obtain satisfactory segmentations without performing OD localization in advance. To overcome this limitation, in this paper, we propose a novel one-stage framework named the multiple graph reasoning network (MGRNet) to segment ODs and OCs from different fundus image datasets. The MGRNet performs graph inference in the Euclidean space, channel space, and hyperbolic space. Euclidean space-based graph reasoning captures the spatial dependencies between pixels from the coordinate dimension. Channel space-based graph reasoning constructs the channel relationship between any pair of channel maps. Hyperbolic space-based graph reasoning provides pixel-level hierarchical embeddings in the hyperbolic space. The MGRNet exhibits obvious advantages over a strong baseline without preprocessing operations, including OD localization, region of interest (ROI) cropping, and data augmentation advantages, and achieves satisfactory results on both the REFUGE (0.9442 mean intersection over union (IoU)) and DRISHTI-GS (0.8792 mean IoU) datasets. Additionally, the MGRNet can also provide reliable glaucoma screening results by measuring the cup-to-disc ratio (CDR) based on segmentation results.

Automated summary

In this paper, a novel one-stage framework called the multiple graph reasoning network (MGRNet) is proposed to segment optic discs (ODs) and optic cups (OCs) from different fundus image datasets. The MGRNet performs graph inference in three different spaces and achieves satisfactory segmentation results. It also provides reliable glaucoma screening results based on cup-to-disc ratio (CDR) measurement.