Article

UPFRONT: Ushering Algorithms Into Retinal Practice

“Have a healthy disregard for the impossible.”—Larry Page, CEO, Alphabet Inc.

In 1996, Stanford University PhD students Larry Page and Sergey Brin started a research project called “BackRub” for the new interactive source of information called the World Wide Web. The key aspect of their project was a technology that helped match pages to a given search string based on checking backlinks to determine a site’s importance instead of how often the search term was used on the page like other search sites of the time. In other words, web pages that were linked from many important pages are themselves also important, and should be ranked higher. The first BackRub webcrawler was housed in Page’s dorm room and he wrote a very basic HTML home page to access the search engine. It became very popular at Stanford and the original program is still housed on Stanford’s servers.

The pair continued to revise their algorithm, which they patented and renamed PageRank. They incorporated their company in 1998, moved it into a garage in Menlo Park, and named it Google. Per Wikipedia, Google now is the most visited website in the world, has an estimated 1 million servers, processes more than 3 billion search requests each day, and evaluates more than 20 petabytes of data each day. Google’s parent company was renamed Alphabet Inc. in 2015, and currently Alphabet is the most valuable public company in the world.

The company’s mission statement is “to organize the world’s information and make it universally accessible and useful,” and toward that end the company has invested heavily in deep learning, where machines use artificial intelligence, neural networks, and scale to crunch large data sets. TensorFlow, its internal deep-learning library, has become the de facto industry standard.

So, what does this have to do with retina? In the future, maybe everything.

Deep-learning computer algorithms could identify diabetic retinopathy with ~90% sensitivity and 98% specificity for detecting referable diabetic retinopathy compared to manual grading by board-certified ophthalmologists in one study. In a Kaggle competition, 650 teams fought for a prize of $100,000 to classify the severity of diabetic retinopathy in 35,000 images. As these algorithms improve, these numbers will surely improve and reading images will be done faster than any clinician. This will allow us to screen thousands of patients and provide care before vision loss. Similar deep-learning initiatives can evaluate OCT images for retina and glaucoma. In this issue, we explore deep learning and retina.

One interesting new app Google is working on is called Trips. The app is a virtual assistant that analyzes your upcoming trips using deep learning to decide what you should see, recommend restaurants, make and manage your reservations, be your local guide, and plan how you get from one place to the next. But what excites me is what is coming from Alphabet’s in-house incubator, ATAP (Advanced Technology and Projects), which is charged with inventing innovative technologies and hardware. While most ATAP projects are secret, they have released information about Project Tango, a positional tracking system that enables mobile devices to map their surroundings without GPS. It’s no wonder that the share value of Alphabet has increased 1,780% since its IPO in 2004.