Glaucoma meets AI

Researchers are employing deep learning to tackle the challenges of glaucoma diagnosis

Default banner image for Glaucoma meets AI
Photo of Cheryl Guttman Krader

[caption id="attachment_15623" align="alignleft" width="1024"] Reflectivity 4.0: a 3D AI companion to simplify and improve glaucoma management[/caption] Image courtesy of Alexandre Thiéry PhD Researchers at the National University of Singapore are aiming to apply deep learning to optical coherence tomography (OCT) images in order to modernise and simplify the diagnosis of glaucoma and to obtain a better understanding of how the disease progresses over time. At the annual meeting of the Association for Research in Vision and Ophthalmology (ARVO) in Vancouver, Canada, Alexandre Thiéry PhD discussed the work being done by the Optical Coherence Tomography & Artificial Intelligence for Glaucomatous Optic Neuropathy (OCTAGON) team. Dr Thiéry is Professor, Department of Statistics and Applied Probability, National University of Singapore, and is co-principal investigator of OCTAGON with Michaël JA Girard PhD, Professor of Biomedical Engineering, National University of Singapore, Singapore. Dr Thiéry explained that the project focuses on the optic nerve head because it represents the main site of glaucoma damage, and it is based on three-dimensional (3D) segmentation. “Glaucoma is a complex disease that affects connective and neural tissues with 3D structural changes. By exploring these 3D structures, we can provide a diagnosis and prognosis of glaucoma more accurately than current gold standards,” Dr Thiéry said. Dr Girard told EuroTimes: “In fact, we have empirically proven that our method is 30% more accurate at diagnosing glaucoma than traditional methods based on fundus imaging or retinal nerve fiber layer thickness assessment.”
Alexandre Thiéry PhD and Michaël JA Girard PhD
BUILDING THE SYSTEM First, the OCTAGON researchers used deep learning to develop an algorithm that could automatically and robustly segment the optic nerve head using a raw OCT scan. “Our method produces expert-level 3D segmentation and is device robust in the sense that it is agnostic to whatever OCT machine is being used for the imaging,” said Dr Thiéry. The software computes in 3D the following unique parameters: prelaminar depth, prelaminar thickness, lamina cribrosa depth, lamina cribrosa curvature, peripapillary scleral angle and choroidal thickness. “People have been focusing on retinal nerve fiber layer thickness as a diagnostic parameter for glaucoma. The main advantage that we see for our approach is that we are able to extract parameters that are highly correlated with glaucoma but that are impossible to measure accurately with conventional methods,” said Dr Thiéry. He added that the algorithm also allows development of interactive software and image enhancement technologies that permits practitioners to annotate the OCTs, explore the optic nerve head and enhance visualisation to ultimately better understand the disease. DEVELOPING THE DIAGNOSTIC TOOL In order to develop an algorithm that could accurately diagnose glaucoma, the team tried initially to use conventional deep learning techniques. After working for several years without success, they turned instead to a different approach. “Conventional artificial intelligence methods are completely unable to capture the complexity of the 3D disease, and more advanced methods were necessary to provide a more accurate diagnosis,” Dr Girard told EuroTimes. First, a neural network was trained with raw OCT volume data in order to segment the neural and connective tissues of the optic nerve head, as well as to extract all of the structural information contained in the OCT volume data. In a second stage, to obtain a robust glaucoma diagnosis, the raw OCT signal, segmentation information and all of the structural 3D parameters were blended into another neural network. Recognising that conventional neural networks do not model uncertainty well, the OCTAGON team also implemented advanced Bayesian methods to achieve more robust probabilistic forecasts. “If you use a conventional neural network to predict whether or not a person has glaucoma, the probabilistic predictions are typically not well calibrated. Indeed, standard deep learning methods are known to not model uncertainty well – that is a big issue when used for medical diagnosis.” Dr Thiéry said. To translate the 3D artificial intelligence technologies developed by OCTAGON for commercial use, Dr Thiéry and Dr Girard founded a start-up company, Abyss Processing, and its 3D artificial intelligence software is now available. “Reflectivity v4.0, which is the latest version, offers unique features for glaucoma diagnosis and prognosis, and it is more powerful than its predecessor as it improves visibility of the optic nerve head in the OCT images through the ability to remove noise, shadows and artefacts,” Dr Girard said. Alexandre H Thiéry: a.h.thiery@nus.edu.sg Michaël JA Girard: mgirard@nus.edu.sg
Tags: optical coherence tomography, Optical Coherence Tomography and Artificial Intelligence for Glaucomatous Optic Neuropathy (OCTAGON)