S-D Net: Joint Segmentation and Diagnosis Revealing the Diagnostic Significance of Using Entire RNFL Thickness in Glaucoma

Jie Wang, Chong Chen, Fei Li, Zhe Wang, Guoxiang Qu, Yu Qiao, Hairong Lv, Xiulan Zhang

Abstract: Glaucoma is a severe eye disease causing blindness. The early diagnosis of glaucoma is of great importance and mainly based on the detection of early signs of optic neuropathy, including retinal nerve fiber layer (RNFL) thinning. We design S-D net to implement automatic segmentation of retinal layers in optical coherence tomography (OCT) images and diagnosis of glaucoma using RNFL thickness vector calculated from the segmentation maps. S-D net is an end-to-end optimized system that mimics the behavior of an ophthalmologist for diagnosing glaucoma with the OCT report. The specially designed unit layer in S-D net gives the threshold of thickness at each point of the RNFL in glaucoma diagnosis. Our results show that S-D net distinguishes glaucoma from healthy cases according to the distribution and magnitude of RNFL thickness. Our model achieves state-of- the-art segmentation results and competitive diagnosing accuracy compared with an experienced ophthalmologist.

Keywords: OCT, segmentation, glaucoma, diagnosis, RNFL thickness