Confidence-Guided Topology-Preserving Layer Segmentation for Optical Coherence Tomography Images With Focus-Column Module

Abstract

Optical coherence tomography (OCT) imaging widely used in retinal examinations yields high-resolution cross-sectional scans of the retina. As a key indicator for studying the development of retinopathy, the change of layer thickness needs to be accurately measured. Although many deep learning-based segmentation methods have been developed, most of them do not explicitly consider the strict order of the retina layers, which easily lead to topological errors. In this article, we propose a novel segmentation framework that employs the distance maps of layer surfaces to convert the segmentation task into multitasking problem for classification and regression, and obtains the topology-guaranteed result through a fusion module. In addition, to cope with the different difficulty in retinal layer segmentation, a confidence network is introduced as a hard-or-easy recognizer to provide online training guidance for the segmentation network. The proposed method is investigated on one OCT data set with slight deformation and two data sets with severe deformation, and the experiment results demonstrate that the proposed method is more effective than other state-of-the-art methods for layer segmentation on OCT B-scans.