As part of Sunday’s “What's New in Ocular Imaging” session (Sym28), Michael F. Chiang, MD, Director of the National Eye Institute, discussed how imaging is the future of artificial intelligence (AI) and introduced some exciting new National Institutes of Health (NIH) research projects driving developments in AI.
Disagreements about diagnoses are a common challenge among clinicians.1 As such, there is strong motivation to use AI, especially in ophthalmology, to improve diagnostic consistency and accuracy. Achieving reliable performance, however, largely relies on the generalizability and usability of the data, and therein lie two key challenges.
Datasets aren’t generalizable. AI systems only work well within the parameters in which they are trained. Referencing a 2019 ROP study,2 Dr. Chiang described how an AI system trained on image data from a North American population and subsequently tested in a similar population showed very high accuracy (AUC: 0.99), but when the same system was applied on a Nepalese population, performance dropped to an AUC of only 0.62. When this particular system was trained using both North American and Nepalese data, however, accuracy of predicting disease in both populations significantly improved (North American AUC: 0.99; Nepalese AUC: 0.98).
Data aren’t easy to extract. A wealth of high-quality data exists, but it is often stored in a variety of devices, which frequently use different data-storage formats and proprietary software. This results in data that are difficult to retrieve or exchange, making it nearly impossible to compile comprehensive, usable AI datasets.
Larger, AI-ready datasets are a necessity. Recognizing that generating larger, nonbiased, AI-ready datasets is the key to being able to effectively use AI, the NIH is currently funding two major projects:
Bridge to AI. The Bridge2AI project is a four-year, $130 million endeavor focused on generating ethically sourced, AI-ready datasets for biomedical and behavioral research communities. By gathering data from a wide variety of sources—including various imaging, clinical, and wearable data—the project aims to improve generalizability and minimize bias while also defining standards for data collection.
AIM-AHEAD. The Artificial Intelligence/Machine Learning Consortium to Advance Health Equity and Researcher Diversity (AIM-AHEAD) project was formed by a consortium of organizations with a mission to address issues in health disparities in minority communities. Beginning with electronic health records, this NIH-wide project seeks to promote the participation of underrepresented researchers and communities and to improve the development of unbiased AI models in health care.
We must insist on ocular imaging standards and collaboration between clinicians and engineers, said Dr. Chiang. Echoing sentiments from the 2022 AAO-ARVO conference that “the use of standard formats for digital imaging is in the best interests of the ophthalmic community and the patients they serve,” Dr. Chiang said, “imaging is powering the revolution in AI,” and ophthalmology is at its forefront.
—Lauren Jarem, MS
Follow Dr. Chiang on Twitter @NEIdirector
1 Campbell JP et al. Ophthalmology. 2016;123:2338-2344.
2 Chen JS et al. Ophthalmol Retina. 2021(5):1027-1035.
Financial disclosures: Michael F. Chiang, MD: None
Disclosure key: C = Consultant/Advisor; E = Employee; EE = Employee, executive role; EO = Owner of company; I = Independent contractor; L = Lecture fees/Speakers bureau; P = Patents/Royalty; PS = Equity/Stock holder, private corporation; S = Grant support; SO = Stock options, public or private corporation; US = Equity/Stock holder, public corporation. For definitions of each category, see aao.org/eyenet/disclosures.
Read more news about Subspecialty Day and AAO 2023.