Detecting Diabetic Retinopathy Using Computer Vision and Deep Learning

Problem Statement

Background

Today, approximately 1 in 10 U.S. adults are currently diagnosed with diabetes according to the CDC. That number is expected to increase exponentially to 1 in 3 U.S. adults by the year 2050[1]. And the U.S. population is only a fraction of the world that is affected by this disease. One of the major complications that can arise with diabetes is an eye disease called diabetic retinopathy. Diabetic Retinopathy (DR) has been one of the leading causes of blindness across the world for many years now. In DR, high blood sugar levels can damage existing blood vessels in the retina by causing swelling and leakage. In addition, abnormal blood vessels often form within the eyes as a result of these conditions[2]. The figure below highlights these differences in the anatomy of the eye when DR becomes dominant. 

Many of these complications often lead to blindness, with almost all patients with Type 1 diabetes and 60% of patients with Type 2 diabetes developing retinopathy during the first 20 years from onset of diabetes[3]. While DR is a very serious condition, it often progresses asymptomatically for many individuals until a sudden loss of vision occurs, making early diagnosis of DR one of the world’s most important and yet difficult leading medical challenges.

Motivation

As mentioned, one of the main concerns with DR is that it often remains undetected until it progresses to an irreversible vision-threatening stage. Currently, DR screening is performed by a retina specialist or trained grader using color fundus photographs (seen in figure below).

However, this leaves a lot of room for error, especially in early stages of DR which are almost undetectable by a human specialist[3]. While DR is a very serious condition that can lead to blindness, progression to vision impairment can be greatly slowed or averted if DR is detected in its early stages.

Additionally, there are many underdeveloped areas of the world where access to the medical expertise and equipment required to detect DR early on is lacking. This is where computational techniques can make a significant impact. For many years, studying the use of computer vision and deep learning methods for detecting early onset of DR has become a major research area. There is a strong need for a comprehensive and automated method of screening for DR using pattern recognition, image classification, and machine learning.

The goal behind this study is to provide a supplemental tool for many parts of the world in helping to prevent vision impairment from DR. For areas that currently have access to expertise and resources, this tool can provide an additional screening factor to help find patterns early on that a human specialist may miss. And in locations where these resources are not available this model can serve as a preliminary tool where a larger number of people can get screened for a low cost at local medical facilities and further pursue medical attention if the screening advises so.