Eyeworld

EW NEWS & OPINION 38 March 2017 by Liz Hillman EyeWorld Staff Writer Research highlight Deep learning algorithm "on par" with ophthalmolo- gists identifying referable diabetic retinopathy and DME T hink of Google, and the search engine, driverless car research, Earth program, or Android operating system is likely to come to mind. The list of technological innovation that Google is involved in goes on, but one slightly less publicized field of its work is in artificial intelligence. Late last year, Google research- ers announced a step forward on this front: they successfully built a machine learning algorithm that was able to learn to detect diabetic retinopathy disease in color retinal images without being engineered to detect specific features of the disease. A study published in the Journal of the American Medical Association in November 2016 described the ability of Google's deep-learning program to learn how to accurately detect di- abetic retinopathy in retinal fundus photos. 1 "We basically showed we are on par with U.S. board-certified ophthalmologists who validated the sets," said Lily Peng, MD, PhD, Google product manager, Mountain View, California. The idea of the program is not to replace eye care professionals, however, but to make the screening process more efficient and available to get those with referable disease into the clinic for professional evalu- ation and treatment. "The primary motivation for us is to improve access to low-cost, high-quality medical diagnostics, and working on a technology that positively impacts everyone," Varun Gulshan, PhD, senior research scientist, Google, said in a statement. According to the World Health Organization (WHO), the incidence of diabetes has risen from 108 mil- lion people in 1980 to 422 million in 2014. 2 Further, WHO states that 2.6% of global blindness is related to diabetes. 3 Just getting diabetes patients to have a yearly dilated eye exam, as recommended by the American Diabetes Association, is a challenge—even in developed coun- tries where access to health care is more readily available compared to developing countries. A study of dia- betes patients 18–64 years old in the United States, for example, found that 59% of white patients and 49% of minority patients received ocular screenings in 2009. 4 "The goal of something like this is to identify patients earlier on in their disease course, at stages that can be intervened at cheaper overall cost to the healthcare system. By getting these individuals in to see the doctor sooner, they can be edu- cated and cared for without having to deal with much more advanced stages of pathology," said Ehsan Ra- himy, MD, vitreoretinal specialist, Palo Alto Medical Foundation, Palo Alto, California, and Google physi- cian consultant. Teaching an algorithm to detect disease A deep convolutional neural net- work was "trained" to detect diabetic retinopathy and diabetic macular edema using a dataset of 128,175 retinal images from the U.S. and three eye hospitals in India. These same retinal images were graded at least three times by ophthalmolo- gists for these conditions, as well as image quality. During the training process, the grade initially guessed by the algorithm was compared to that de- termined by the ophthalmologists, with the function being adjusted accordingly to decrease its error. According to Gulshan et al., this process was repeated for each image several times and in doing so "the function 'learns' how to accurately compute the diabetic retinopathy severity from the pixel intensities of the image for all images in the train- ing set." The study authors added that while the algorithm does not detect lesions specifically, it seems to recognize them as local features. From there, the trained algo- rithm was validated against two other datasets, which were graded by board-certified ophthalmologists. Even at very high sensitivity—the accuracy of the algorithm to identi- fy people with disease—it was also found to have high specificity, not calling out false positives, explained Peter Karth, MD, Stanford Univer- sity and Oregon Eye Consultants, Eugene, Oregon, a vitreoretinal specialist also serving as a consultant and grader in the study. Although other deep learning programs have exhibited high sensitivity, Dr. Karth said, not all also had such success with high specificity. "[T]he present study extends this body of work by using deep convolutional neural networks and a large data set with multiple grades per image to generate an algorithm with 97.5% sensitivity and 93.4% specificity," Gulshan et al. wrote. "When screening populations with substantial disease, achieving both high sensitivity and high specificity is critical to minimize both false-pos- itive and false-negative results." In addition to showing that such a system can be trained and clinically validated for this pur- pose, the study authors conducted further experiments with training and tuning of the algorithm. They concluded that to adequately train a deep learning program for medical purposes, the initial dataset needs to be large with tens of thousands of abnormal images. The tuning, or how accurate the system is, plateaued in this case with about 60,000 images. Currently, the algorithm has only demonstrated its ability to identify diabetic retinopathy and diabetic macular edema, but Drs. Rahimy and Karth envisioned further refinement to include de- tection of other sight-threatening conditions as well, such as tumors, age-related macular degeneration, and glaucoma. "There is a lot more work to be done, but the first hurdle they cleared is very exciting because it does validate the technology," Dr. Rahimy said. Limitations to the algorithm, acknowledged by the authors in the study, include that the reference standard was a majority decision of the ophthalmologist graders; the exact features being used by the neural network to make predictions is unknown; and possibly the design of the influence of the online system on the performance of ophthalmol- ogists grading images. "Machines will never replace doctors" Drs. Rahimy and Karth emphasized how this technology is not going to replace physicians but actually serve to increase services to patients who need help. "There is always going to be a role for physicians; machines cannot replace doctors. What they're going to be able to do is augment the care we provide. It's supposed to be syn- ergistic," Dr. Rahimy said. "Through having this artificial intelligence platform, physicians can expect growth in their practices by expand- ing the number of patients seeking and requiring treatment. There are Google's artificial intelligence program detects diabetic retinopathy

• Blank Page
• Blank Page (1)