In a study sponsored by Google, researchers found that a newly developed deep learning computer algorithm taught itself to detect diabetic eye disease as well as 54 medical eye specialists (ophthalmologists) when reviewing photos of the retina.
Deep learning is a computational method which allows an algorithm to program itself through “learning.” The project’s system, called DeepMind, “learns” by studying a large set of examples that demonstrate the desired behavior and then adapting itself in response.
The algorithm and the eye specialists diagnosed cases of diabetic retinopathy and diabetic macular edema from retinal fundus photographs, such as those shown below.
In reviewing the data, one of the DeepMind researchers, Google’s Lily Peng, MD, PhD, observed, “The results show that our algorithm’s performance is on-par with that of ophthalmologists. For example, on the first validation set, the algorithm has an F-score
What is the algorithm’s place in a clinical setting? Right now, researchers say it cannot replace a human eye doctor, because it has only learned to recognize signs of diabetic retinopathy and diabetic macular edema.
“It may miss non-diabetic retinopathy lesions that it was not trained to identify,” Dr. Peng noted. “Hence, this algorithm is not a replacement for a comprehensive eye exam, as it will ignore components such as visual acuity, refraction, eye pressure measurements, etc. However, with further research, the results suggest that the algorithm could lead to improved care and outcomes compared with the current ophthalmologic assessment.”
This new deep-learning algorithm has potential use in telemedicine, as it will allow patients to “self-diagnose” in the comfort of their own homes, even if it is only for diabetic retinopathy and diabetic macular edema. “Along with telemedicine, technologies such as these could increase access to care and assist in screening for diabetic retinopathy in areas where there are few eye doctors,” Dr. Peng claimed.
Dr. Peng noted that deep learning algorithms may reveal insights not previously recognized by eye specialists.
“Because the network ‘learned’ the features that were most predictive, it is indeed possible that the algorithm is using features previously unknown to or ignored by humans. Understanding which features are used to make predictions will be an important area of investigation for further studies, and is indeed an active area of research within the larger machine-learning community,” Dr. Peng said.