Following is a summary of this Mid-Year Forum 2018 session examining the application of artificial intelligence in ophthalmology.
The next transformation in ophthalmology is the application of artificial intelligence in diagnosing and treating disease in clinical practice. It is beginning to be used in retinal disease for detecting diabetic retinopathy and diabetic macular edema from fundus photographs, and its potential is to provide more efficient and objective analysis of images and prediction of disease progression. We will explore what artificial intelligence means for practicing ophthalmologists, its promise and limitations, and what the future holds.
Rahul Khurana, MD, noted that artificial intelligence is neither a good nor evil robot of the future but rather a technology of today and is being utilized all around us. Artificial intelligence was developed in the 1950s, but it has become popular today due to advances in computing power to making it feasible, and because of the explosion of “Big Data.” Many groups have turned to artificial intelligence to manage big data effectively. Artificial intelligence is defined as the engineering of making intelligent machines and programs. Machine learning is the ability of machines to learn without being explicitly programmed. Deep learning allows machines to train themselves to perform tasks by exposing multilayered neuronal networks to vast amounts of data. Neuronal networks typically have hundreds to thousands of layers and the processes on how to come up with answers have been described as black box. Artificial intelligence is especially effective for image recognition; and is already being used in many specialties in medicine such as radiology, dermatology, pathology, and now ophthalmology. In April, the first artificial intelligence system was approved by the FDA to detect diabetic retinopathy, so this topic is of great relevance to ophthalmology.
Summary of Comments from Guest Speakers
Application of Artificial Intelligence in Retinopathy of Prematurity
Michael F. Chiang, MD, professor of ophthalmology and medical informatics and clinical epidemiology and vice-chair (Research), Oregon Health and Science University Casey Eye Institute
Our goal as ophthalmologists is to provide the best care. However, there is an expanding rate of knowledge, and a lengthy lag between significant discovery and adoption into routine care, as evidenced by a study showing that 50% of patients with acute myocardial infarction received beta-blocker medications 16 years after the beta-blocker heart attack clinical trial. Retinopathy of prematurity (ROP) is the leading cause of childhood blindness and is treatable if diagnosed early in life. The evaluation of infants is performed using indirect ophthalmology with hand-drawn sketches for documentation. The challenges are that performing examinations are difficult, impressions are imprecise and subjective, and the number of ophthalmologists performing these exams is small because of a high medicolegal liability, and the time and costs for examining these patients in the hospital. There is an international standard for disease staging, vascular parameters and terminology, (International Classification of Retinopathy of Prematurity – ICROP) providing an infrastructure for clinical trials and clinical practice. Plus disease is the most critical parameter for severe treatment-requiring ROP, with the key features being arterial tortuosity and venous dilation upon examination that can be exemplified in a standard published photograph. The goal of artificial intelligence is to help clinicians more accurately diagnose by quantifying the vascular parameters with automated image analysis. This is accomplished with accurate segmentation of vessels and validation against a robust reference standard that has already been developed by the profession. Two strategies could be used to address this question of whether we can reduce the variability and increase the accuracy of ROP diagnosis: classic machine learning, or deep learning using convolutional neural networks (CNN). A study trained CNN for ROP used 6,000 posterior pole images and the findings were positive, resulting in an Area Under Curve (AUC) of 0.98 to identify plus disease. In conclusion, ophthalmic diagnosis is inherently subjective and qualitative, for example, the tortuosity of vessels in ROP, and neovascularization in diabetic retinopathy (DR). This leads to significant inconsistency, even among leading experts in the field, and this indicates that the performance of “real-world” physicians may be even more inconsistent. Thus, in these settings, there is a potential role for expert systems to help physicians improve the accuracy and consistency of diagnosis. In ophthalmology, we are fortunate to have ongoing development and validation of these technologies, collaborative and interdisciplinary research, and large data sets for evaluation. But we can’t lose sight of the role of clinical judgment on the part of the physician and how best to integrate that art and science of medicine.
Deep Learning for the Detection of Diabetic Eye Disease
Lily Peng, MD, PhD, product manager, Google
Dr. Peng described her current project at Google on using deep learning for detecting DR. Diabetic retinopathy is the fastest growing cause of blindness and is estimated to total up to 415 million people worldwide. Regular screening is the key to preventing blindness. There is a need to improve the screening of patients with DR, particularly in areas where patients don’t have access to ophthalmologist care. The requirements for machine learning are well-delineated standards for diagnosis, a large dataset for training and tuning and other datasets for validation. A few years ago, Google embarked on training a deep neural network to read fundus images with 130,000 fundus images and using other datasets for validation with another 10,000 plus fundus images. This initiative resulted in a study published in JAMA in 2016 that demonstrated an AUC of 99.1 for the diagnosis of DR or DME warranting a referral to the ophthalmologist. The next steps to enhance this model of machine learning are to use better reference standards and address explainability. The better reference standards being considered are finer grading schemes such as the Early Treatment of Diabetic Retinopathy Study (ETDRS) scale, and adjudication with retinal specialists, OCT imaging for diabetic macular edema and ultrawide field imaging for DR. A study published in Ophthalmology in 2018 found a 0.84 weighted kappa using the ETDRS scale with the machine learning. Regarding explainability, neural networks are not really a black box, but use logical learning approaches to gain knowledge, and future evaluation will identify what we cannot explain now through human observation alone. This could then lead to completely new, novel scientific discoveries. This has been demonstrated in this technology’s prediction of age, sex and cardiovascular risk factors in study published in Nature in 2017. The next steps are to expand into other disease areas and other imaging modalities for this exciting technology.
Artificial Intelligence: What's Really Next for Healthcare
Krishna Yeshwant, MD, MBA, general partner, GV
Dr. Yeshwant predicted that the future will not herald advances in this area as in the field of artificial intelligence but as tools to better diagnose and manage disease, and ultimately be seen more as a routine part of patient care, which would be the right impact of artificial intelligence. To contextualize artificial intelligence, in the 1950s, the question was, can create artificial human intelligence in a computer? We are still nowhere near this. Instead, we are at the stage of intelligence augmentation, taking what it is to be human and augmenting with additional tools. One tool is the neural network using a propagation of algorithms, mimicking what humans do but in a more automated and quicker way. Autonomous searching is an augmentation of human memory. Automated image recognition is an augmentation of human diagnostic skills. Intelligence augmentation in medicine is moving in profoundly exciting areas. Intelligence infrastructure is not just a propagation of algorithms, but the synthesis of different inputs from devices, electronic health record systems, images, etc. to make better diagnoses and treatment decisions. On a societal scale, this translates to a learning health system, gaining knowledge from data flowing from patients, testing instruments, genomes, studies from the scientific literature, and pulling all the information and interactions back into the 1:1 patient and physician relationship. Physicians can utilize machine learning to ask relevant questions and respond to the need to provide care in a more accessible and affordable manner. Large corporations are active in this space, but professional groups will need to be involved and will influence its adoption. Ultimately, artificial intelligence will complement the skills of physicians, provide a new skillset to the exam room and enhance the compact of physicians to society to provide better care for their patients.
Summary of Audience Comments
- Will ophthalmologists be replaced by artificial intelligence? Artificial intelligence will instead, overlay clinical practice; instead, physicians will be relied upon to collate and integrate artificial intelligence to make sound decisions. This is the integration of the art of medicine with the science of medicine. The goal of these systems is to provide a tool to improve the accuracy and consistency of diagnosis, but not to replace physicians. This highlights the importance of maintaining uniquely human skills, and the blending of the art and science of medicine.
- Will there be transparency in the motives and choices behind artificial intelligence? Dr. Peng explained that the choice to evaluate diabetic retinopathy through deep learning was based upon the existence of standards for diagnosis, and the large impact on populations around the world. This made it ideal for the model case but it’s likely that deep learning will be applied to other diagnoses within ophthalmology. To ensure that inequities of care are addressed, a larger diversity of investigators in deep learning is needed.
High Priority Objectives
- To keep current with different artificial intelligence activities, including deep learning, machine learning, and provide ophthalmological input and guidance.
- To provide an understanding about how these artificial intelligence activities can augment and improve ophthalmologists’ tools to diagnose and manage patients, and how ophthalmologists can integrate these tools in their day-to-day clinical practice.
Review more sessions in the Mid-Year Forum 2018 Report.