Hossein Nazari Khanamiri, MD
Dr. Bernerd Doft, president of the Retina Society, opened the first Innovation in Retina lecture series dedicated to the application of artificial Intelligence (AI) and neural networks in retinal disease. He challenged all audience members to think about how we can harness these innovative technologies in our practices and how our clinics will evolve when such technologies become widely available.
Dr. Dimitri Azar, former dean of University of Illinois College of Medicine and now the senior director of ophthalmology at Verily, gave his fascinating talk titled, “The Future is Now: Artificial Intelligence in Ophthalmology and Retina.” He started with an introduction to neural network and deep learning. Dr. Azar then elaborated on the various methods utilized for machine learning and the current applications of artificial intelligence in ophthalmology that include artificial intelligence (AI) utilized to stratify corneal topographies before refractive surgeries, using AI for the diagnosis of dry eye and glaucoma, and analyzing fundus photographies and OCT scans. These technologies may soon be available for AMD screening and predicting AMD progression. Finally, he noted the current limitations of applying AI to ophthalmology including arbitrary diagnostic criteria of various disease conditions that are considered as “gold standard” for AI based diagnostic algorithms, image quality, and detecting and correction of errors.
Then, Dr. Taiji Sakamoto from Kagoshima University, Japan, presented data on machine learning-based quantification of choroidal vessels on en-face images in central serous chorioretinopathy (CSCR). Dr. Sakamoto enlightened the audiences about developing and refining the machine learning algorithms, which showed that vessels in Haller’s layer of the choroid were dilated and less symmetrical vertically in eyes with CSCR.
Dr. Neil Bressler from Wilmer Eye Institute then gave his captivating talk about using deep learning techniques for stratifying retinal images based on the AREDS 9-step detailed severity scale. Deep learning grading technique was promising in grading the severity of images based on a set of criteria that normally require highly trained graders. This method could potentially provide an automated and accurate estimate of 5-year risk of progressing to advanced AMD.
Later, Dr. Michael Singer from Medical Center Ophthalmology Associates discussed using artificial intelligence to distinguish different retinal pathologies. Utilizing a deep learning algorithm called “transfer learning network” allowed the training of a highly accurate model with a relatively small training dataset. Using this technique on fundus images and OCT scans, diabetic macular edema, choroidal neovascularization, drusen, and normal retina was accurately differentiated with a high degree of sensitivity and specificity.
Next speaker, Dr. Dilraj Grewal from Duke Eye Center, reported the development and performance of a deep convolutional neural network (DCNN) program for the diagnosis of birdshot chorioretinitis. After successfully training the DCCN on the color fundus photographs, Dr. Grewal’s group showed that artificial intelligence assisted diagnosis yielded an accuracy of 96.8% for distinguishing birdshot chorioretinitis from other diagnoses.
Dr. Saad Shaikh gave the last talk of this session about the automatic detection and quantification of retinal fluid in OCT images, also utilizing the DCNN architecture. After training the network, the system achieved promising detection sensitivity per volume of 0.94, 0.92, and 1.00 for intraretinal fluid, subretinal fluid, and pigment epithelial detachment, respectively.
As the investigators have demonstrated, artificial intelligence is no longer in the realm of science fiction. The technology has positioned itself to likely become an integral part of our profession – and perhaps may touch many aspects of our lives. While further developments, studies and validations are required prior to widespread use, next year’s Retina Society meeting will likely host even more talks about leveraging artificial intelligence technologies to potentially further improve care for our patients.