JOINT LEARNING ANNOTATOR CALIBRATION AND ANNOTATOR PREFERENCE FOR MULTIPLE ANNOTATIONS OPTIC DISC AND CUP SEGMENTATION

The accurate and automatic segmentation of optic disc (OD) and optic cup (OC) from fundus images plays an important role in the screening and diagnosis of glaucoma. However, optic disc and cup annotations suffers from annotator variation due to the inherent differences in annotators' expertise and the inherent blurriness of retinal fundus images. In clinical practice, considering the opinions of multiple annotators can effectively reduce the impact of this annotator-related bias. In this paper, we propose an efficient framework to joint learn annotator calibration and annotator preference for multiple annotations optic disc and cup segmentation, which consists of two main parts. In the first part, we model multi-annotation as a multi-class segmentation problem to learn calibration segmentation. Further, we employ a cascaded architecture that introduces the anatomical knowledge of the optic disc and optic cup, which can effectively improve the segmentation performance of the optic cup. In the second part, each annotator's preference-involved segmentation is estimated through annotator encoding and conditional convolution learning. Experiments on the RIGA benchmark show that our framework outperforms a range of state-of-the-art (SOTA) multi-annotation segmentation methods.