Artificial Intelligence for Automated Detection of Diabetic Retinopathy

Nitish Mehta, MD
Ophthalmology Resident
New York University

Ehsan Rahimy, MD
Surgical and Medical Vitreoretinal Specialist
Palo Alto Medical Foundation

In April 2018, the FDA granted marketing approval to IDx for the first medical device utilizing artificial intelligence to detect referable diabetic retinopathy (DR) from color fundus photography. Targeted to primary care providers, the cloud-based software, IDx-DR, analyzes images obtained from a Topcon NW400 fundus cameras and recommends referral to an eye care professional if the image satisfies criteria for greater than mild diabetic retinopathy. IDx-DR was granted ‘Breakthrough Device’ designation by the FDA and is the first device authorized for marketing that provides a screening decision without the input of a clinician.

An exciting (and sometimes anxiety-provoking) technology, artificial intelligence in its various forms is rapidly integrating itself into the framework of modern society. In medicine, the combination of a growing volume of electronic clinical data and a rapid increase in computing power is allowing for the development of algorithms that hold promise to reshape and disrupt the practice of clinical medicine. Ophthalmology, and in particular the field of Retina, is particularly primed for the application of artificial intelligence due to a heavy reliance on imaging, an expanding role for tele-medicine, and an overall shortage of specialists available to handle a growing and aging patient population. For example, an estimated 50% of people with diabetes do not get annual eye exams.1 It is suggested that accurate identification of referral-warranted disease by automated retinal image analysis systems may efficiently enhance the ability of patients to access appropriate care.


Artificial Intelligence

Terms such as machine learning, artificial intelligence, and deep learning are becoming engrained in the popular vernacular. Generally, artificial intelligence is defined as the ability of computer systems to perform complex, independent tasks that require human-like intelligence such as visual processing, speech recognition, or decision-making. Machine learning is employed when computer programs have the ability to improve their own decision making by ‘learning’ from data provided to them without provided explicit rules. Deep learning, an increasingly popular and powerful model of machine learning, utilizes layers upon layers of ‘neural networks’ to enhance the software’s ability to independently perform feature extraction from data.

Deep Learning

The architecture of deep learning software is roughly modeled on the human brain. The first step involves providing the algorithm training data such as a large collection of images pre-labeled by Diabetic Retinopathy Severity Scale (DRSS). Each node, or neuron, performs a mathematical operation, such as the identification of a certain pattern of pixels, on the input data. Neurons in parallel, a convolutional neural network (CNN), then release their result to another neural network which then performs a similar set of calculations. Ultimately, the entire network releases an output, and the classification of the image is compared to the original input. Each node’s calculation carries a certain weight, and the weights are continually adjusted by the software until classification error is minimized (training). Rather than writing code to identify microaneurysm, for example, the algorithm develops a set of ‘hidden’ rules to learn how to categorize images on its own. This is one of the most intriguing applications of deep learning. The ‘black box’ rules are currently hidden to the developers, and if utilized correctly could perhaps provide new insights into our own data.


IDx-DR’s approval was based on a clinical study that assessed the software’s performance on retinal images from 900 patients with diabetes at 10 primary care sites. IDx-DR was able to correctly identify the presence of greater than mild DR 87.4% of the time and was able to correctly identify those patients did not 89.5% of the time. Using a reading center derived-ETDRS score of greater than 43 from stereo-images of the same patients as a benchmark, IDx-DR’s sensitivity and specificity for detecting greater than mild DR was 87% and 90%, respectively. The company reports that the system is easy to use: inexperienced operators who received a one-time standardized four-hour training program were able to image patients and transfer information to the platform 96% of the time. Patient with any active visual complaints or with a certain subset of prior ophthalmic history are not considered appropriate for screening with IDx-DR.


A Growing Space

IDx LLC is one a several players in a rapidly growing space. Other automated retinal image analysis systems include iGradingM (Medalytix Group Ltd, Manchester, UK), Retmarker (Retmarker SA, Taveiro, Portugal), and EyeArt (Eyenuk, Woodland Hills, California). In the past few years, a variety of publications have reported implementation of deep learning on fundus photos and OCT images obtained from patients with diabetic retinopathy and age-related macular degeneration. As the platforms themselves utilize different proprietary software and training sets, direct comparison is difficult and reported accuracy statistics vary from study to study.

Breaking the Black Box

The use of publicly available datasets such as the Kaggle EyePACS or Messider-2 may allow for a more direct comparison of accuracy and reliability. Some platforms have the ability to ‘highlight’ regions of interest on an image that the algorithm deemed noteworthy, enhancing its transparency. Ultimately, close-observation of a real-world application of the platforms will be required.

The potentials of artificial intelligence in retinal imaging is immense, including identifying new imaging biomarkers and guiding treatment decisions. How best to utilize this technology will be an important consideration for the future of retinal care.
1. Murchison AP, Hark L, Pizzi LT, et al. Non-adherence to eye care in people with diabetes. BMJ Open Diabetes Res Care 2017;5(1):e000333.