Diabetic retinopathy (DR) is one of the leading causes of blindness in the world, affecting up to 80% of people who have had diabetes for 20 years or more. It is usually asymptomatic until it has progressed to advanced stages and affects the patient's vision (See Fig 1 below). To prevent DR, all diabetics are advised to get their eyes screened once a year. However, only 15% follow-up on this advice, largely due to lack of access to eye specialists and screening equipment.
Recently there has been great progress in building AI models for detecting DR, fueled by the increased research on artificial neural networks and deep learning. An artificial neural network is a structure for modelling and learning from data, where the basic building blocks - the artificial neurons, are inspired by neurons in the brain. Neural networks are the core of many of the most advanced models within artificial intelligence, as they have shown superior performance on a number of modelling tasks and on different kinds of data. They are trained using examples of inputs and the corresponding output we want the network to produce, and since they often contain millions of parameters, they require very large datasets for training. The models for diagnosing diabetic retinopathy are usually trained to output the grade of severity of diabetic retinopathy given a retina image, and they are trained using retinal images graded by experts.
In 2018, IDx-DR became the first AI based software for detecting DR to be approved by the FDA, and is now being used in clinics to give patients with diabetes instant feedback on the progression of the disease, without the need for an eye specialist. The same year, a research team in Google AI showed that their DR models achieved kappa scores close to the scores reached by retinal specialists, meaning that they were performing close to human expert level on this task. The joint efforts made by research teams and companies around the world to develop more secure, accurate and robust models will make screening for diabetic retinopathy accessible to a huge number of people that previously didn’t have access to eye specialists.
Although artificial neural networks have existed for a long time, the wide-spread use of them in research and development started only recently, and there are still many unanswered questions about their stability and robustness. The huge dimensionality of neural networks makes it hard to truly understand what is going on inside and what the networks actually learn, and many use them as a black box. The models stability and accuracy are extremely critical in medical applications, such as software for diagnosis, because a false result could potentially have very serious consequences for the health of the patient.
Oivi’s mission is to create an analysis platform that will provide more insight into the progression of the disease, and create transparency around how the diagnosis is set. This platform will help the primary care physician or a specialist in diagnosing the stage of DR and deciding on the right treatment at the right time. We are excited to take part in the movement of bringing AI into the physicians office, so that early diagnosis will become accessible and affordable for large populations, helping doctors make better and more informed decisions.
Author: Ragnhild Holden Helland
Machine Learning Engineer at Oivi
This article and its contents are owned by Oivi.