Commentary: Rise of machine learning and artificial intelligence in ophthalmology

Abstract

Less than a decade ago, artificial intelligence (AI) and machine learning were more a part of science fiction, and were often depicted as going rogue in several fictional movies. Skynet in “Terminator”, HAL9000 in “2001: A Space Odyssey”, Ultron in “Avengers: Age of Ultron”, the sentient machines in “The Matrix”, Sonny in “I, Robot”, and the supercomputer in WarGames are just a few examples. But with the coming of the Fourth Industrial Revolution, AI has crept up stealthily into our daily lives. This industrial era is characterized by a fusion of technologies referred to as cyberphysical systems. It is marked by revolutionary technology innovations in fields like artificial intelligence, machine learning, 3D printing, robotics, industrial Internet of Things, autonomous vehicles, and so on. There is artificial intelligence in our smartphone digital assistants (Google Now, Siri, Alexa, Cortana, Bixby), Gmail (email filters, smart replies, reminders), Facebook (newsfeed, image recognition, proactive detection), Amazon (product recommendations), Maps (route planning with traffic data), chatbots, and much more. Dermatology, radiology, pathology, and ophthalmology are leading the AI wave in healthcare, due to the large volume of images to process to obtain a diagnosis. Ophthalmology is a very visual subject, and there are a lot of images that we have to see and analyze, including fundus images, retinal SD-OCT (spectral domain optical coherence tomography), RNFL (retinal nerve fibre layer) OCT, anterior segment images, slit images for AC depth, AS-OCT, corneal topography, visual field perimetry, Hess charting, diplopia charting, 9 gaze images, A-scan, B-scan, and a few more. These lend themselves to the possibility of image processing and analysis using artificial intelligence and machine learning. Deep learning, which uses convolutional neural networks (CNNs) is a subset of machine learning, which itself can be considered to be a subset of AI. Way back in 2016, Google had reported the use of a deep CNN to create an algorithm for automated detection of diabetic retinopathy (DR) and diabetic macular edema in retinal fundus photographs.[1] Although it had a high sensitivity (97.5%) and specificity (93.4%), there was the Black Box problem, which meant that the AI could not explain what features in the images it had used in the CNN to arrive at the diagnosis. However, in April 2019, Google has published how they used Integrated Gradients Explanation to show a heatmap on the fundus image to show the features the deep CNN used to make the diagnosis. They showed that this opening up of the Black Box improved the accuracy of, and confidence in, DR grading in an AI-assisted grading setting.[2] The main areas of ophthalmology where major strides in AI[3] have been made are in analyzing fundus images of DR, age-related macular degeneration (ARMD), retinopathy of prematurity (ROP), retinal vein occlusion (RVO), and glaucoma.[4] Some work has also been published about grading cataract, analyzing topography, predicting progression of myopia, and detecting ocular surface squamous neoplasia (OSSN) from unstained histopathology specimens. The author is currently working on AI algorithms related to glaucoma and has seen a few commercial AI algorithms do their work and they have great potential, to say the least. Rajalakshmi et al. published in March 2018, their study on automated AI screening of fundus photos taken on an iPhone using Remidio Fundus on Phone (FOP) showing high sensitivity (95.8%) and specificity (80.2%) for detecting DR. These machine learning algorithms are getting better at diagnosis, leading to April 2018, when IDx-DR became the first FDA-approved AI software to screen fundus photos for DR. These software (or their innovative variants)[5] can potentially run on any smartphone, which can be converted into a smartphone fundus camera such as DIYretCAM[6] or T3Retcam[7] for less than a dollar. The accompanying review article, titled “Ar