The Ophthalmic Artificial Intelligence Summit, held on June 18, 2022, included didactic presentations and discussions from renowned retina specialists, this summit aims to focus specifically on how AI can impact the practice of ophthalmology today.
Sessions 1 & 2
Usha Chakravarthy, MD, PhD, CBE, discussed whether fluctuations in macular fluid volumes are clinically significant, and how AI analytics can help clinicians better understand the impact of these fluctuations on visual outcomes. Results from studies evaluating automated AI applications in a real world data set were presented, and showed that visual outcomes are strongly influenced by the location of the fluctuating fluid. A second study evaluating fluid fluctuations with the The Port Delivery System with Ranibizumab compared to Ranibizumab monthly injections, showed that visual acuity outcomes were similar across arms, regardless of fluctuations. She concluded by noting that fluid in the retinal environment in nAMD is an abnormal state of affairs, and that intraretinal fluid is the most concerning, highlighting that control of disease activity without recurrences can help maintain retinal architecture and function.
Emily Chew, MD reviewed AI algorithms for AMD as informed by the AREDS and AREDS2 Studies, using deep learning to evaluate multiple features of AMD, including drusen size and area, reticular pseudo-drusen, hyperpigmentary or hypopigmentary changes, and late AMD (neovascular or geographic atrophy). In almost all instances, deep learning was superior to the human gradings for detection of drusen (including size) and pigmentary changes. However, for late AMD, deep learning was not superior to the human gradings. Deep learning models predicting risk of AMD using images from AREDS were also discussed, demonstrating that survival analyses can predict with high accuracy progression to late AMD, either nAMD or GA, using different algorithms and including multi-modal and multi-task algorithms.
Sobha Sivaprasad, MS Ophth, DNB, DM, FRCS(Ed), FRCOphth discussed using artificial intelligence to identify non-responders post aflibercept loading phase for neovascular AMD. Noting that approximately half of patients receiving monthly anti-VEGF treatment have evidence of persistent fluid after 2 years, the objective of the analyses presented was to predict the presence of residual fluid 8 weeks after initiation of treatment, using baseline OCT characteristics. Results of the analyses showed that age and central subfield thickness correlated with persistent fluid, and the ensemble AI system successfully predicted the treatment response, highlighting the potential of using AI to identify appropriate patient groups prior to entering a trial or initiating a treatment assignment.
Sessions 3 & 4
Ursula Schmidt-Erfurth, MD presented latest developments in the use of AI tools for identification of disease activity and therapeutic response in patients with geographic atrophy. AI-based segementation of OCT volumes allowed automatic detection of photoreceptor (PR) loss and alteration of the morphology of the PR layer. In addition, the analysis of the PR layer was used to show that patients receiving monthly treatment with pegcetacoplan had a reduction in the progression of PR loss and thinning. Interestingly, the effect of therapy is more pronounced at the PR level which precedes and exceeds RPE loss in GA disease, highlighting the possibility of using AI-based OCT analysis to predict GA progression and therapy effect.
Michael F. Chiang, MD presented the perspective from the National Eye Institute on how support for AI research can be increased. He noted that some of the outstanding unanswered questions regarding the use of AI include generalizability and bias, noting that results from one population can’t be extrapolated to other populations. Another challenge is the diagnostic variability, with unclear definitions of clinically-significant endpoints. Some of the potential solutions discussed were data sharing and harmonization, establishing standard data representations as well as standards for image collection. Two recent projects were also discussed, highlighting the incentives to improve the creating and utilization of datasets, such as Bridge2AI and AIM-AHEAD.
Pearse A. Keane, MD provided an update on the collaboration between Moorfields Eye Hospital and DeepMind, highlighting the latest developments from this productive collaboration. The algorithm developed by the collaboration can segment different types of PED, and segments the RPE as well, which allows identification of drusen and areas of GA. In addition, the segmentation and classification network allows for the algorithm to be used in multiple OCT systems. Highlighting the complexity of using AI algorithms, it was noted that retina specialists are still not using AI routinely, and that one of the main reasons for that is the need to bridge the AI chasm, which is the difference between using AI for research studies vs regulated by clinicians. A number of studies are currently ongoing to emphasize the clinical validation of AI tools, and the future is bright for these approaches!
Sessions 5 & 6
J. Peter Campbell, MD, MPH reviewed applications of AI in Retinopathy of Prematurity (ROP), highlighting that AI may have very relevant applications in the diagnosis, monitoring, and management of the disease. The goal of using AI in ROP is to not only have an objective diagnosis of the disease, as most of the time patients do not have any symptoms, but also monitor disease progression, to identify in advance when patients progress to needing treatment and using insights to decide how patients should be treated (e.g. laser vs anti-VEGF). Importantly, AI may be used to reduce the number of examinations needed to identify high risk patients and could reduce screening burden by 50%.
Minhaj Nur Alam, PhD, presented a contrastive learning pipeline for robust diabetic retinopathy classification, with the objective of improving the management for retinopathies and enabling mass screening, early onset detection, and clinical diagnostics. The FundusNet model was used to evaluate and compare referrable vs non-referrable retina images, reducing annotation efforts by 90%, and the model generalized well. Some of the learnings shared included the observation that augmentation and transformation are key to better representation learning, and that the model trained better with high batch size.
David Myung, MD, PhD discussed incorporating AI into a diabetic retinopathy screening workflow in an academic primary health care network, highlighting real world experience in implementing AI into practice. He reviewed how Stanford has implemented a Telemedicine network and workflow, using the HEDIS measure which provides a number of key measures for patients with diabetes, including retinal imaging. Using the system IDx-DR, which is an FDA-authorized autonomous AI device for DR screening in the primary care setting, diagnosis of DR with a human hybrid workflow can be done at multiple primary care sites. Of note, implementation of AI into DR screening increased adherence with diabetic eye exams, improving quality of care and expanding access to necessary eye care for diabetic patients.
Yasha Modi, MD discussed a machine learning model for predicting outcomes and treatment frequency following monthly aflibercept for macular edema secondary to central retinal vein occlusion (ME-CRVO). The datasets used for the study were from the Copernicus and Galileo studies, and included patients who switched to PRN treatment and subsequently had BCVA and fluid fluctuations. The goal was to use the data to develop a machine learning model to predict change in BCVA and change in CST, as well as to help manage patients” and physicians’ expectations during management of ME-CRVO. The model was able to successfully predict visual outcomes, change in macular thickness, and dosing frequency with accuracy. In addition, it was shown that CST at baseline and at week 4 were two key factors confirmed to predict PRN dosing frequency after fixed dosing.
Sally L. Baxter, MD, MSc provided an overview of the use of AI and predictive modeling of glaucoma progression, using electronic health records (EHR) data from the NIH All of Us Research Program, a large and diverse nationwide database. Noting that the digitalization of health data has provided an opportunity to better understand complex relationships between systemic disease and glaucoma, the goal of the research was to determine whether data-driven modeling of systemic attributes obtained from existing EHR data predict the risk of glaucoma progression. Predictive modeling using All of Us offered a large number of participants and greater diversity, allowing models trained with this data to achieve superior predictive performance than single-center models.
Darius Moshfeghi, MD presented a photographic ROP severity score for treatment decision making, called PROP-SS, noting that it includes core features of ICROP-Zone, Stage, Plus, it can be used to create a simple photographic ROP severity score for acute phase screening. A prospective validation of the model is currently ongoing, emphasizing the need for a simplified and single score system that will be extremely useful in ensuring eligible babies are screened and scored.