recognition specialist are numbered. Turns out, machines are really good at pattern recognition—better than you and I.
In these pages we provide a crash course on deep learning with additional background on page 21 thanks to William Ou. The general concept of machine learning is that one teaches a machine to detect patterns in data through application of rules, accomplished by providing explicit definitions of features to consider in drawing a conclusion. Such technology has been used for decades in the interpretation of electrocardiograms,2 producing the printout at the top that most of us rely on to determine if there’s reason for concern.
Deep learning represents a subset of machine learning in which no external rules are provided. The algorithm programs itself by learning from a large collection of examples that illustrate a defined finding. The
algorithm learns the most predictive findings directly from the images themselves. Such deep-learning systems are already commercially employed for radiologic applications.
They are coming to retina. Google sponsored and published a deep-learning algorithm for detection of “referable” diabetic retinopathy and “referable” diabetic macular edema based on training from a dataset of over 250,000 retinal images. Notably, just two of the 15 authors are MDs.3
Fascinatingly, despite our ability to create and implement deep-learning algorithms, we have limited insight into the specific features employed to arrive at a given conclusion. How these algorithms “see and think” are a black box and a field of intensive investigation.
“We are doctors,” you say, “not computer scientists or mathematicians!” Get ready.