Deep Learning Algorithms for Glaucoma Diagnosis based on Optical Coherence Topography

Abstract

Objectives: Developing and validating deep-learning algorithms to differentiate optic coherence tomography (OCT) images from subjects with and without glaucoma, by analysing their retinal nerve fibre layer (RNFL) morphologies. Methods: Optical coherence tomography (OCT) was performed on subjects. Images of their peripapillary retinal nerve fibre layer (RNFL) was used for analysis. A two-step process involving the use of cascaded deep learning algorithms was developed for glaucoma diagnosis. The first step was to train a U-net deep learning algorithm on 152 randomly selected cross-sectional optical coherence tomography images, to segment the RNFL. The resulting algorithm was then used to segment RNFL from 70 glaucoma patients and 90 normal subjects. In the second step, the segmented images were fed into another deep learning classification algorithm for differentiation between glaucoma patients and normal subjects. Cross-validation was then employed to evaluate performance of the algorithm over unknown datasets. Results: An accuracy of 99% was achieved for the RFNL segmentation, with an average dice coefficient of 0.99 on a validation dataset, while an area under curve of 96% was observed for the classification algorithm. Conclusion: Deep-learning algorithms for OCT optic nerve head images analysis in an effective and accurate appraoch to assist in arriving diagnosis of glaucoma, given that the morphological changes as a result of the condition, such as RFNL thining, are sufficiently isolated.