A Multi-device, Multi-ethnicity Deep Learning Algorithm to Detect Glaucoma from A Single Optical Coherence Tomography Scan of the Optic Nerve Head

Abstract

Purpose : To develop a multi-ethnicity and multi-device deep learning algorithm that can detect glaucoma from a single 3D optical coherence tomography (OCT) scan of the optic nerve head (ONH).

Methods : Diagonal & Raster OCT scans were acquired from 4 OCT machines (Spectralis, Cirrus, Atlantis, Nidek RS-3000) through the center of the ONH for subjects from Australia (Sydney), China (Hong Kong, Bei Jing and Zhong Shan), Lithuania (Vilnius), Russia (Ufa) and Singapore. For consistency, each OCT scan was then converted to a radial scan with 6 B-scans passing through the center of Bruch’s membrane opening. In total, 55,902 B-scans were used from 5,747 normal subjects and 39,096 B-scans from 2,363 glaucoma subjects (Figure 1.a). For each OCT device, all scans were balanced through up-sampling, and split into a training set (70% of scans), a validation set (15%) and an independent testing set (15%). To classify glaucoma from non-glaucoma subjects, we developed a custom convolutional neural network (deep learning) that exploited 3D structural information about neural and connective tissues of the ONH. Extensive online data augmentations (e.g. translation, rotation, flipping, and noise addition) were performed to improve accuracy. We calculated the area under the receiver operating characteristic curve (AUC) to assess whether our deep learning algorithm was able to differentiate glaucoma from non-glaucoma eyes.

Results : Our deep learning algorithm was able to accurately classify glaucoma from non-glaucoma with an AUC of 0.90 (Figure 1.b) on the independent testing set.

Conclusions : We propose a novel deep learning approach that provides a glaucoma diagnosis from a single OCT scan, using multi-device and multi-ethnicity datasets. Our approach may simplify glaucoma diagnosis as it is fast (diagnosis is computed in less than 1 second), and it only requires 1 clinical test (i.e. OCT imaging of the ONH).

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.