Spatial Pyramid Dynamic Graph Convolution Assisted Two-Stage U-Net for Retinal Layer and Optic Disc Segmentation in OCT Images

Abstract

Retinal nerve fiber layer (RNFL) thickness in retinal optical coherence tomography (OCT) images is commonly used in the diagnosis of glaucoma. However, due to the presence of the optic disc, the retinal tissue surrounding the optic disc is difficult to segment. To solve this problem, this paper uses a two-stage U-Net as the inference framework, inserts a pyramid dynamic graph inference module in the two-stage U-Net framework, and performs coarse-to-fine graph feature inference between the encoder and the decoder. Finally, a two-stage segmentation model SpDGRU-Net is proposed to segment the retinal layer and optic disc respectively. This paper conducts experiments on the OCT public dataset, and the proposed SpDGRU-Net segmentation network achieves an average Dice score of 0.826 and an average pixel accuracy of 0.835, both of which outperform other state-of-the-art techniques.