Eyeworld

14 Ophthalmology Business • September 2018 by John D. Banja, PhD Artificial intelligence in ophthalmology: Clinical practice and business implications 29% of persons with diabetes in the US.—the prevalence of congenital cataracts (CCs) is very low, with a projected incidence of 4.39 per 10,000 children in the U.S. (com- pared to 2.2–13.6 per 10,000 children in various countries throughout the world). 6 Nevertheless, an estimated 200,000 children throughout the world are bilaterally blind from congenital cataracts, making it the primary cause of treatable childhood blindness in the world. Somewhat like Gulshan's screening study, Qian Wang and Dinggang Shen did a study interfacing a cloud-based platform with an AI system trained to screen for congenital cataracts. 7 They used a "convolutional neural network" (CNN) that was developed by Yizhi Liu and colleagues to identify CC patients, which it did with about a 99% accuracy, then suggest treatment options. But since ophthalmologic services, especially for rare diseases, are scarce in many parts of the world, Wang and Shen proceeded to inter- face the screening system with cloud computing that could refer patients from various sites to the best region- al care available. In their system, registered users could upload ocular images from a remote site, have the CNN assist with the diagnostic and treatment plan, and using Wang and Shen's platform, have those patients referred to collaborating hospitals that were equipped to provide the necessary services. The benefits of such technology are obvious, not only given AI's diagnostic accuracy for a rare condition, but in the expe- ditious way the best ophthalmologic services can become available. tions. In 2016 Varun Gulshan and his colleagues published a study in JAMA that used AI to screen persons for diabetic retinopathy using retinal fundus photographs. 3 The researchers trained an algorithm from 128,175 previously rated images for referable diabetic retinopathy, diabetic macu- lar edema, and overall image quality. The algorithm was then tested on 9,963 images from 4,997 patients that had been graded by at least seven U.S. board-certified ophthal- mologists. The software exhibited an area under the curve for detecting referable diabetic retinopathy of 0.99. Commenting on the study, Andrew Beam and Isaac Kohane noted that this technology, which relies on a special type of computer chip, can easily be implemented into existing computer systems at a cost of around $1,000. 4 The technology is able to process about 3,000 images per sec- ond, which translates to 260 million images per day (because the device can work non-stop). In yet another commentary on Gulshan's work, Tien Yin Wong and Neil Bressler specu- lated that this kind of technology would greatly expand ophthalmo- logic services in underserved areas, where diabetic retinopathy reading centers could be set up. 5 Images from remote sites could be transmitted to such centers, where an initial screening would be done and referral recommendations made and relayed back. The clinical contribution this technology could make to the world- wide prevalence of diabetic retinopa- thy seems incalculable. While the prevalence of diabetic retinopathy is high—affecting about A safe bet is that artificial intelligence (AI) will continue to improve its various functionalities and, as the years go on, will alter the land- scape of human productivity, quality of life, and global business practices in ways that are presently unimag- inable. IBM's Watson, which beat the best Jeopardy players in history in 2011, can analyze 200 million pages of text in 3 seconds, illustrating the jaw-dropping computational power of these devices at storing, compre- hending, analyzing and retrieving facts and knowledge from very large databases. 1 That capacity has direct applications to healthcare, as "big data" storage and analysis—for ex- ample, the immense amounts of data that can be gleaned from millions of medical records and images—will enable diagnostic, prognostic, and treatment models that will increas- ingly become the norm and, indeed, the standard of care, over the next 2 decades. AI is already amazingly good and probably will become astonish- ingly good at image recognition and classification—so much so that it has sparked considerable worry over the future of image-reliant specialties like radiology and pathology. 2 I wondered about the applications of these tech- nologies in ophthalmology, specifi- cally with a view to how they might change ophthalmology workflow and certain aspects of its business model. I came across two fairly recent studies that illustrate the promise of AI in ophthalmology but that also raise interesting marketplace ques-

• Blank Page
• Blank Page (1)