DEEP LEARNING APPROACHES ON OPTICAL COHERENCE TOMOGRAPHY IMAGES FOR IMAGE RESTORATION, SEGMENTATION, AND GLAUCOMA DIAGNOSIS

Abstract

In this work, we aimed to develop deep learning (DL) approaches to maximize diagnosis accuracy of glaucoma by restoring and segmenting optical coherence tomog- raphy B-scans. First, we proposed a novel technique that can restore information within retinal blood vessel shadows. Second, we presented a DL algorithm to remove speckle noise. This technique also reduced the chance of motion artifacts from mani- festing in OCT B-scans. We show that this technique preserves the integrity of the B-scan and removes speckle noise and retinal blood vessel shadows simultaneously with a single pass of a single network. We showed that the developed algorithm could improve segmentation performance on B-scans created by a different OCT manufacturer. Thirdly, we developed a novel method to diagnose glaucoma from the 3D structure of retinal blood vessels.