Use of Artificial Intelligence in Monitoring Age-Related Macular Degeneration
By Tian Xia, MD; Marco A. Zarbin, MD, PhD; and Neelakshi Bhagat, MD, MPH
The prevalence of age-related macular degeneration (AMD), a major cause of blindness in industrialized nations worldwide, is increasing. AMD is projected to affect 196 million persons in 2020 and 288 million by 2040.1 It is estimated that in the United States, by 2030, one in five individuals over the age of 65 years will have AMD.2 Given current therapeutic options, early detection and prompt treatment are critical for maintaining as much vision as possible. Artificial intelligence (AI) is an emerging field, and its application in retinal screening could have an important impact on the management of patients with AMD.
Most applications of AI are using machine learning (ML), where learning is achieved through training examples; the computer identifies a pattern and generates algorithms and then tests known data to check the accuracy of the algorithms and recalibrates them until the output matches the expected result. A specialized type of ML, deep learning (DL), is increasingly being used in retinal imaging. It uses artificial neural networks to builds algorithms similar to how the human brain works. These networks comprise a large number of processing units called nodes (neurons) that operate in parallel and are interconnected by links. Each link is associated with weights that contain information about the input signal. Each node has an internal state of its own, termed activation level, which is a function of the inputs that the node receives (analogous to the way the axon hillock summates dendritic input from numerous synapses). Mathematical algorithms are used to define the activation function that generates output from these summated inputs as well as to define the matrices that connect the nodes. The recursive rearrangement of computational architecture is "machine learning"2 and involves modifying the weights in the connections between network layers with the goal of achieving an expected (correct) output. Mathematical algorithms define the training paradigm where the learning process (e.g., supervised, unsupervised and reinforcement learning) requires detailed entry of labelled training images of the disease and a validation data set. Convolutional neural networks (CNNs) involve convolution (mathematical combination of two functions to produce a third function) for each feature of interest (e.g., in a fundus photo), subsampling (to smooth inputs from different convolutional layers), activation (to control how signals flow from one layer to the next because some output signals activate more nodes than others) and full connectedness (in which the last layers of the network are fully connected, which mimics high level reasoning where all possible paths from input to output are contemplated).
AI was first developed to perform automated image analysis through simple pattern recognition with basic algorithms and now has evolved to advanced deep learning using CNNs.3,4 "DL" uses previously labeled image data to train computer programs to detect disease and even grade disease severity. To achieve the most accurate model, the machine can self-modify the algorithm until it achieves the one that produces results to closely match the real data.
Several studies highlight the plausibility of AMD screening through AI.3,5,6 Early AI studies used clinical features depicted on fundus images such as drusen, retinal pigment epithelial detachment (RPED), etc. as markers to diagnose AMD and its severity. The diagnostic sensitivity for AMD in these studies ranged from 75% to 100%.7 Some algorithms can predict progression of dry AMD by quantifying drusen over time.8 Currently, most researchers are using spectral domain optical coherence tomography (SD-OCT) image markers to detect progression of AMD; these studies have shown high sensitivity and specificity.3,6,9 Treder et al, using TensorFlow software through deep convolutional neural network (DCNN) learning, created a model from a training set of 1,012 horizontal cross-sectional SD-OCT images of normal posterior poles as well as images of exudative AMD. After an iteration of 500 training steps, the computer could detect AMD accurately in 100% of cases. Using the DCNN classifiers created in the last step of training, an untrained set of 100 SD-OCT images was tested with 100% sensitivity, 92% specificity and 96% accuracy in detecting the disease.9
Various automated approaches to diagnose and discern progression of AMD have been tried.10-13 The AI algorithms for exudative macular degeneration have focused on detection of retinal fluid.8 Bogunovic et al have utilized an automated model to observe the response to anti-vascular endothelial growth factor (anti-VEGF) treatment using OCT images.14 Schmidt-Erfurth et al evaluated the enrollment SD-OCT images of patients in the HARBOR trial.15 Using supervised machine learning regression to predict best-corrected visual acuity at 12 months based on extracted morphologic imaging biomarkers (various retinal layer thickness, area and volume of intraretinal fluid, subretinal fluid and pigment epithelial detachment), the machine predicted visual acuity at 12 months post treatment with 71% accuracy.15 The most predictive image biomarker was horizontal extension of intraretinal cystoid fluid in the foveal region.10,14 Other pilot studies have used AI algorithms to evaluate RPED, geography atrophy (GA) and drusen quantification through contour-based segmentation of RPED and GA. Although AI is in its infancy for detection, progression and response to anti-VEGF therapy, progress is being made rapidly.
Regarding monitoring for development or recurrence of choroidal neovascularization (CNV) in patients with AMD, the ForeseeHome monitoring device (Notal Vision Ltd.) has been approved by the U.S. Food and Drug Administration since 2009. This machine uses preferential hyperacuity perimetry testing to detect progression of metamorphopsia, a marker directly related to worsening of disease or CNV growth. The HOME study used this detection device with automated results that were reported directly to the clinician’s office. Patients using home monitoring had less visual acuity loss at the time of CNV detection compared to patients with standard monitoring with a median loss of 4 letters compared to 9 letters.16 Moreover, the HOME study report 3 disclosed that the home device strategy was associated with a higher detection rate compared to standard office visiting monitoring with a relative risk of 16.0.17
Technology assistance for management of AMD has shown promise in early detection of disease as well as monitoring of at-risk eyes. However, it is not without flaws. AI can only be as good as the quality of the training sets. With some deep machine learning based on self-generated rules, there is the inherent problem of machine errors based on clinically insignificant image changes. Nonetheless, with more technological advances, there will likely be a greater presence of AI in the field of ophthalmology for telemedicine screening and disease management.
Wong WL, Su X, Li X, et al. Global prevalence of age-related macular degeneration and disease burden projection for 2020 and 2040: a systematic review and meta-analysis. Lancet Glob Health. 2014;2(2):e106-e116.
Lee A, Taylor P, Kalpathy-Cramer J, Tufail A. Machine Learning Has Arrived! Ophthalmology. 2017;124(12):1726-1728.
Burlina P, Joshi N, Pacheco KD, Freund DE, Kong J, Bressler NM. Utility of Deep Learning Methods for Referability Classification of Age-Related Macular Degeneration. JAMA Ophthalmol. 2018 Nov 1;136(11)1305-1307.
Burlina PM, Joshi N, Pekala M, Pacheco KD, Freund DE, Bressler NM. Automated Grading of Age-Related Macular Degeneration From Color Fundus Images Using Deep Convolutional Neural Networks. JAMA Ophthalmol. 2017;135(11):1170-1176.
Ting DSW, Cheung CY, Lim G, et al. Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations With Diabetes. JAMA. 2017;318(22):2211-2223.
Fang L, Cunefare D, Wang C, Guymer RH, Li S, Farsiu S. Automatic segmentation of nine retinal layer boundaries in OCT images of non-exudative AMD patients using deep learning and graph search. Biomed Opt Express. 2017;8(5):2732-2744.
Mookiah MR, Acharya UR, Fujita H, et al. Local configuration pattern features for age-related macular degeneration characterization and classification. Comput Biol Med. 2015;63:208-218.
Schlanitz FG, Baumann B, Kundi M, et al. Drusen volume development over time and its relevance to the course of age-related macular degeneration. Br J Ophthalmol. 2017;101(2):198-203.
Treder M, Lauermann JL, Eter N. Automated detection of exudative age-related macular degeneration in spectral domain optical coherence tomography using deep learning. Graefes Arch Clin Exp Ophthalmol. 2018;256(2):259-265.
Schmidt-Erfurth U, Waldstein SM. A paradigm shift in imaging biomarkers in neovascular age-related macular degeneration. Prog Retin Eye Res. 2016;50:1-24.
Güven A. Automatic detection of age-related macular degeneration pathologies in retinal fundus images. Comput Methods Biomech Biomed Engin. 2013;16(4):425-434.
Mora AD, Vieira PM, Manivannan A, Fonseca JM. Automated drusen detection in retinal images using analytical modelling algorithms. Biomed Eng Online. 2011;10:59.
Chen Q, Leng T, Zheng L, et al. Automated drusen segmentation and quantification in SD-OCT images. Med Image Anal. 2013;17(8):1058-1072.
Bogunovic H, Waldstein SM, Schlegl T, et al. Prediction of Anti-VEGF Treatment Requirements in Neovascular AMD Using a Machine Learning Approach. Invest Ophthalmol Vis Sci. 2017;58(7):3240-3248.
Schmidt-Erfurth U, Bogunovic H, Sadeghipour A, et al. Machine Learning to Analyze the Prognostic Value of Current Imaging Biomarkers in Neovascular Age-Related Macular Degeneration. Ophthalmology Retina. January 2018;2(1):24-30.
Chew EY, Clemons TE, et al; AREDS2-HOME Study Research Group. Randomized trial of a home monitoring system for early detection of choroidal neovascularization home monitoring of the Eye (HOME) study. Ophthalmology. 2014;121(2):535-544.
Chew EY, Clemons TE, Harrington M, et al; AREDS-HOME Study Research Group. Effectiveness of Different Monitoring Modalities in the Detection of Neovascular Age-Related Macular Degeneration: The Home Study, Report Number 3. Retina. 2016;36(8):1542-1547.
About our author(s):
Dr. Xia is an ophthalmology resident (PGY-4) at Rutgers New Jersey Medical School and is interested in pursuing a career in vitreoretinal surgery.
Dr. Zarbin is the chair of Institute of Ophthalmology and Visual Science at Rutgers New Jersey Medical School.
Dr. Bhagat is professor of Ophthalmology and director of Vitreoretinal Surgery at Rutgers, New Jersey Medical School.