TAU: Transferable Attention U-Net for optic disc and cup segmentation

Abstract

Glaucoma is the leading cause of blindness globally. It causes irreversible damage to the optic nerve fibers and leads to vision deterioration. Automated methods for glaucoma detection have been proposed to intervene in an early stage and reduce the lesion. The core modules of the prevalent methods are typically supervised segmentation networks, trained with annotated eye fundus images. The outputs of these models are the segmentation maps of the optic disc (OD) and optic cup (OC), which are used for pathological analysis. However, eye fundus images captured by different fundus cameras vary significantly in resolution, contrast, and field-of-view. Models pretrained on one dataset suffer performance degradation on other datasets owing to domain shift. Moreover, limited by the high cost of manual annotations, it is not practical to annotate all the sources of eye fundus images. In this paper, we propose a transferable attention U-Net (TAU) model for OD and OC segmentation tasks across different eye fundus image datasets. Two discriminators and attention modules are applied in the proposed model to locate and extract invariant features across datasets. Experiments on three widely used datasets confirm that TAU outperforms four baseline methods in terms of segmentation performance and cup-to-disc ratio-based glaucoma detection accuracy.