Louis Cai, MD
Ophthalmology Resident
Wills Eye Hospital
There has been an explosion of interest in artificial intelligence (AI) as clinicians become more familiar with the technology and its applications. Dr. Suber Huang moderated an excellent section reviewing some of the latest work done in this field for ophthalmology.
Dr. Aaron Lee opened the session with his study, Multicenter, Head-to-Head, Real-World Validation Study of Seven Automated Artificial Intelligence Diabetic Retinopathy Screening Systems. Working with five different companies (ADCIS, AirDoc, Eyenuk, Retina-AI Health, Retmaker), Dr. Lee tested the performance of seven different algorithms to screen 311,604 images for diabetic retinopathy. He extracted fundus photos from both the Puget Sound VA in Seattle and the Atlanta VA, and used the original VA teleretinal grades as the “truth.” The images from the Atlanta VA were also graded by clinicians to compare and validate the AI algorithms. The patients included in this study were predominantly male (94.7%) with Type II Diabetes Mellitus (100%). Patients were routinely dilated in the Atlanta VA, resulting in a significantly lower proportion of ungradable images (2.5%) when compared to those taken at the Seattle VA (16.2%), where dilation was less routine. All algorithms exhibited high negative predictive value, but low positive predictive value. Compared to the human grader, two algorithms had higher sensitivity but lower specificity for identifying any DR, and one algorithm was indistinguishable from clinician grading. In regards to detecting moderate NPDR or worse – the clinician grader had a 100% sensitivity and was matched by three algorithms. Dr. Lee did not find algorithmic biases with race, but did note that specificity seemed to be worse in patients with older age. Given the excellent overall performance and limited biases of these particular algorithms, Dr. Lee believes they are ready for deployment. His work demonstrates the necessity for AI algorithms to be validated on real world data prior to adoption in practice.
Dr. Tien Yin Wong followed with an informative talk titled Prediction of Systemic Diseases From Eye Images Using AI and Deep Learning. He provided a comprehensive review of some of the recent studies that used color fundus photos to predict systemic diseases. As Dr. Wong explained, fundus photos are an ideal platform for screening as they are cheap, easily taken, and non-invasive. Additionally, there are established clinical guidelines for many diseases including hypertension and diabetes, and large annotated datasets already exist. Although AI has found success in estimating certain systemic risk factors like age, blood pressure, and BMI, the majority of systemic biomarkers are likely unable to be derived from color fundus photos alone. Dr. Wong then presented studies that show the potential for AI to replace existing biomarkers, citing a study from Chang et al. (2020) using color fundus photos to predict the severity of atherosclerosis and subsequent cardiovascular disease as well as a study by Rim et al (2021) using retinal photos to estimate coronary artery calcium (CAC), a cardiovascular biomarker typically measured with CT scans. Narrower retinal arterioles and wider retinal venules have been associated with negative cardiovascular outcomes, and deep learning can utilize these biomarkers to assess cardiovascular disease. Oculomics is the field of study using ocular biomarkers to detect systemic disease. AI has found success in detecting diseases directly, including anemia, chronic kidney disease, and type 2 diabetes. The hurdles to adopting AI in practice are non-technical including fear of the “black box” approach, clinical process changes, regulatory/reimbursement hurdles, and varying support from non-ophthalmologist clinical champions.
Dr. Ursula Schmidt-Erfurth then presented her work on AI-Based Fluid Monitoring In Clinical Practice. She opened by describing how OCT has revolutionized the way we make clinical decisions, especially for patients requiring chronic anti-VEGF injections. AI can expedite and improve clinical decision-making using OCT images, especially as the number of patients with AMD and the number of OCT images captured continues to grow. I based clinical decision support systems have smaller error rates compared to humans. AI can detect fluid, localize it, and quantify fluid volumes, which is lacking in conventional OCT analysis. Currently, prospective studies are using automated quantification of fluid volumes to guide treatment decisions. Measuring dynamic fluctuations of fluid can provide more accurate predictions of BCVA loss, as higher fluid volumes (IRF, SRF, PED) are associated with vision loss. Overall, these automated methodologies can reduce human resource burdens and provide objective treatment parameters for both clinical trials and real-world practice.
Finally, Dr. Pearse A Keane finished the session with Clinician-Driven Machine Learning: A New Phase for AI-Enabled Health Care?. He described code-free deep learning, which he sees as the next phase of AI-enabled healthcare, allowing clinicians to develop AI models with minimal coding experience. His colleagues were able to download datasets of skin cancer lesions, chest X-rays, OCTs, and fundus photos and utilize code-free deep learning techniques to create state-of-the-art AI algorithms. Multiple code-free deep learning platforms exist, including those created by Apple, Amazon, Clarifai, Google, MedicMind, and Microsoft, and all have demonstrated relatively good performance for the classification tasks tested.
Dr. Keane replicated the results of landmark AI studies using code-free deep learning. One surprising study used color fundus photos to identify patient gender – not typically a classification task that retinal specialists can train themselves to perform. Ultimately, code-free deep learning can democratize AI for clinicians, allowing these techniques to be available to all – not just those with computer science expertise.