Building an Open AI-Powered Ophthalmic Healthcare Innovation Ecosystem: Challenges, Opportunities, and Strategic Roadmap
In recent years, artificial intelligence (AI) technology has advanced rapidly, emerging as one of the forefront research hotspots in the medical field. Notably, owing to its convenience and efficiency, AI demonstrates significant application potential in the screening, diagnosis, treatment, and follow-up care of ophthalmic diseases.
In the field of ophthalmic AI, research in China is on par with world-class standards. So, what value can the application of artificial intelligence technology in our country bring to solving difficult problems in ophthalmic medical care? What challenges exist in building an AI ecosystem for ophthalmic healthcare? What are the trends in the development of AI in ophthalmic medicine? I am honored to have invited Professor Zhang Xiulan, Council Member of the Asia-Pacific Glaucoma Society, Fellow of the Asia Pacific Academy of Ophthalmology, and Director of the Clinical Research Center at Zhongshan Ophthalmic Center, to provide an in-depth interpretation of relevant issues in the field of ophthalmic AI. It is hoped that this will offer insights to readers concerned about the development of medical artificial intelligence.
In developed Western countries, there are an average of 79 ophthalmologists per million people. China currently has approximately 35,000 ophthalmologists, averaging only 22 per million people. Moreover, there is a significant disparity in the professional competency of ophthalmologists across different regions and hospitals of varying tiers. Furthermore, with population aging, the number of patients with eye diseases continues to rise. The training cycle for ophthalmologists is exceptionally long, requiring at least 7–12 years to become a qualified practitioner. Therefore, merely increasing investment in ophthalmologist training is insufficient to meet the current demand for diagnosis and treatment of eye diseases.
China has the largest population of blind and low-vision individuals, as well as the highest number of patients with eye diseases globally, imposing a heavy burden on the nation, society, and families.
Taking cataracts as an example, the relevant surgical techniques are highly mature. In China, the number of cataract surgeries per million people has increased from 83 in 1988 to 2,205 in 2017; however, this figure still falls far short of the rate in Europe and the United States, where more than 10,000 procedures are performed annually per million population.
Glaucoma is the leading cause of irreversible blindness in China. It is projected that by 2020, there will be 21 million glaucoma patients in China, resulting in nearly 6.3 million blind individuals and over 10 million people with visual impairments. The onset of most glaucoma cases is insidious; at least 90% of primary open-angle glaucoma cases and at least 50% of primary angle-closure glaucoma cases in China remain undiagnosed.
Diabetic retinopathy is another serious public health issue in China. Currently, there are 13 million patients with diabetic retinopathy in the country, and diabetic patients in rural areas face a higher risk of developing this condition than their urban counterparts. If detected early, vision-threatening complications can be effectively prevented through laser surgery; however, screening for diabetic retinopathy remains infrequent among diabetic patients in China. A survey in Guangdong Province revealed that 43.2% of diabetic patients had never undergone an ophthalmic examination, with this figure reaching as high as 81.1% in rural areas.
The primary pain point in the diagnosis and treatment of eye diseases is the severe mismatch between the objective demand for clinical care—particularly for screening—and the limited number of ophthalmologists.
“Clinical Practice Guidelines for Diabetic Retinopathy in China (2014)” recommend that patients with diabetes undergo screening for retinopathy at least once a year. With over 100 million individuals affected by diabetes in China, relying solely on ophthalmologists is insufficient to meet the substantial screening demand. Due to the specialized nature of ophthalmology and the scarcity of qualified professionals, diagnosis and treatment coverage remains extremely limited. However, the integration of artificial intelligence with large-scale imaging data can facilitate screening and diagnosis for certain ocular conditions, such as diabetic retinopathy, cataracts, glaucoma, and macular degeneration.
From a physician’s perspective, there are inherent limitations to the diagnosis and treatment of human diseases, and early diagnosis of many conditions remains challenging. Taking glaucoma as an example, it is difficult for ophthalmologists to determine whether the disease is progressing based on a single outpatient visit. Similarly, in patients with diabetic retinopathy receiving anti-VEGF therapy, individual treatment responses cannot be predicted beforehand; instead, comprehensive assessments through multiple follow-up visits are required.
Artificial intelligence (AI) technology may assist ophthalmologists in making early assessments in this regard, thereby facilitating personalized diagnosis and treatment. Furthermore, although human hands are dexterous, they have inherent limits in precision. By leveraging AI technology to develop ophthalmic surgical robots that assist in eye surgeries, their high precision, flexibility, and stability can be closely integrated with minimally invasive techniques, potentially enabling remote ophthalmic surgery in the future. Currently, AI applications for diagnosis and decision-making represent a focal point of AI research, whereas the development of AI-powered surgical robots remains less mature than diagnostic AI.
Overall, AI can assist ophthalmologists in the following five areas:
1. By applying machine learning technologies integrated with medical imaging in ophthalmology, AI can assist physicians in screening, significantly expanding the coverage of diagnosis and treatment; this enables early detection of high-risk or affected populations through screening, facilitating timely intervention.
2. AI can assist in clinical diagnosis, improve the diagnostic efficiency of ophthalmic diseases in clinical practice, and alleviate the workload of ophthalmologists; these two aspects are achievable in the near term and are urgently needed in our country, as they can help to some extent address the challenges of difficult access to medical care and a shortage of doctors, thereby revolutionizing the existing disease diagnosis and treatment system.
3. AI can expand the capabilities of ophthalmologists, guiding personalized treatment and predicting prognosis. Of course, achieving this relies on the construction of longitudinal datasets.
4. Assisting in the training of ophthalmologists. For instance, various AI-integrated simulators are currently used for resident training; existing AI products can also help primary care physicians compare their diagnoses, thereby facilitating their professional development.
5. AI can simulate human reasoning to synthesize novel diagnostic approaches and patterns, thereby assisting physicians in clinical decision-making and improving diagnostic accuracy. However, achieving this remains challenging and depends on further technological advancements.
Professor Zhang Xiulan believes that AI algorithms based on fundus photography are now relatively mature, with diagnostic capabilities for fundus and optic nerve diseases approaching the level of human physicians, and demonstrating certain efficacy in screening for common ophthalmic conditions. However, there remain objective challenges to the clinical implementation of AI in ophthalmology.
Figure: Professor Zhang Xiulan, Council Member of the Asia-Pacific Glaucoma Society, Fellow of the Asia Pacific Academy of Ophthalmology, and Director of the Clinical Research Center at Zhongshan Ophthalmic Center
From a technical perspective, the performance of algorithms in precisely localizing specific lesions still needs improvement. This is primarily due to the current lack of large-scale, finely annotated imaging databases. Furthermore, the interpretability of these algorithms remains low, with the so-called “black box” problem persisting, which limits their widespread adoption. In addition, there is currently no real-world evidence on the performance of AI algorithms; therefore, their practical utility and reliability when applied to general populations remain unknown.
Even so, some vendors, such as Baidu, have made significant progress in interpretability. They have integrated single diagnostic tasks performed by black-box models into the overall architecture design and developed interpretable algorithms that align with clinicians’ reasoning logic, based on clinical diagnosis and treatment pathways.
From a non-technical perspective, numerous challenges remain to be addressed. AI in healthcare in China is still in its nascent stage, with official standards currently being drafted; as of now, there is no unified standard for data annotation. Although China possesses vast amounts of medical imaging data—far exceeding those available overseas—the quality of this data is uneven. Labels and annotations are often rudimentary, and significant clinical information is missing, which severely compromises algorithm performance. While the “volume” of data is important, its “quality” is even more critical. In the future, those who possess high-quality data will emerge as the winners in the AI landscape.
Meanwhile, the government and industry associations need to establish quality monitoring standards, formulate approval and regulatory standards for artificial intelligence products, and conduct standardized evaluations for each AI product. Public acceptance of AI-powered medical services will take time; the business models for AI in healthcare also require further clarification. In terms of ethics and law, it remains to be determined who bears responsibility for misdiagnoses and medical malpractice involving AI.
At present, there are no ophthalmic AI products that have received certification from the China Food and Drug Administration (CFDA). The field of intelligent screening for fundus diseases has seen the earliest breakthroughs, with numerous smart screening products for conditions such as diabetic retinopathy and age-related macular degeneration (AMD) currently in the clinical validation phase, though none have yet been launched on the market. However, given the abundance of imaging examinations in ophthalmology, the field is poised for a flourishing era of diverse ophthalmic AI innovations. As a pioneer in domestic ophthalmic AI, Zhongshan Ophthalmic Center led the establishment of China’s first Ophthalmic Intelligence Group last year, with Professor Zhang Xiulan participating in the initiative. The group aims to promote the development and growth of ophthalmic AI in China and will also contribute to drafting standards for medical data annotation.
In the field of ophthalmic AI, research in China is currently on par with world-class standards. The Zhongshan Ophthalmic Center of Sun Yat-sen University is a pioneer in domestic ophthalmic AI, having achieved industry-renowned results in medical AI research related to congenital cataracts, diabetic retinopathy, glaucoma, and myopia.
In the realm of academic research, in 2017, Professor Liu Yizhi, Director of the Zhongshan Ophthalmic Center, spearheaded the development and construction of the world’s first artificial intelligence cloud platform for cataract diagnosis and treatment. This work was published as a cover article in Nature Biomedical Engineering. Additionally, the center launched the world’s first AI-driven ophthalmology outpatient clinic in Guangzhou. As the only project completed by a Chinese team to be selected, it was recognized by IEEE Spectrum as one of the “11 AI Events That Influenced the Global Medical Community.”
Based on big data, President Liu Yizhi demonstrated the real-world patterns of onset, progression, and stabilization of myopia in adolescents, thereby establishing an artificial intelligence prediction system that enables personalized forecasting of myopia progression trends. Furthermore, the research team at the Zhongshan Ophthalmic Center, Sun Yat-sen University, participated in international multicenter studies on AI algorithms for fundus photography, developing algorithms based on color fundus images for screening diseases such as diabetic retinopathy and age-related macular degeneration (AMD). Related articles have been published in top-tier general medical journal JAMA, leading diabetes journal Diabetes Care, and Ophthalmology, the number-one ranked journal in ophthalmology.
The team led by Professor Zhang Xiulan focuses on research into intelligent diagnostic algorithms for glaucoma. Through close collaboration with the Shenzhen Institute of Advanced Technology of the Chinese Academy of Sciences, Baidu, and more than ten ophthalmic institutions in China, the team has developed iGlaucoma 1.0, a visual field interpretation system based on deep neural networks, to distinguish between glaucomatous and non-glaucomatous visual fields. Based on results from tens of thousands of labeled visual field reports, the team has published papers reporting a diagnostic accuracy of 87.6%, and multi-center clinical validation is currently underway.
The team is collaborating with 10 centers in China to jointly develop iGlaucoma 2.0, a neural network capable of simultaneously interpreting multimodal reports comprising visual field tests and optical coherence tomography (OCT). Further research will leverage labeled data from anterior segment OCT, visual field tests, and posterior segment OCT to develop artificial intelligence algorithms, thereby establishing iGlaucoma 3.0, an AI-based glaucoma diagnostic platform that integrates both anterior and posterior segment imaging data. Due to the complexity of glaucoma diagnosis, the development of relevant AI algorithms presents significant challenges, making this a long-term endeavor. The team has formulated a comprehensive follow-up research plan with the goal of creating a multimodal AI decision-support platform for glaucoma diagnosis, ultimately aiming to assist clinicians in clinical decision-making.
Just as Professor Fei-Fei Li from the Department of Computer Science at Stanford University sparked the current wave of artificial intelligence research through the development of the ImageNet database, data is equally critical in the field of medical AI. Currently, there are no unified official standards for ophthalmic AI; however, it is widely acknowledged that involving more annotators and accumulating larger datasets generally result in higher-quality databases. Since database construction directly impacts algorithm development and performance enhancement, it constitutes a crucial component of the process.
Due to the lack of comprehensive databases to support training for lesion localization, current ophthalmic AI is limited to diagnosis, with relatively weak capabilities in precise lesion localization. Professor Zhang Xiulan and her partners recognize that the current bottleneck in medical AI development stems from the scarcity of high-quality imaging data.
Therefore, since 2018, the Zhongshan Ophthalmic Center has begun to engage in the construction of public databases. In September 2018, the Zhongshan Ophthalmic Center, together with Baidu and the Vienna Eye Hospital, released the world’s first fully open, meticulously annotated fundus photography database for glaucoma at MICCAI, the premier conference on medical imaging.（1）, attracting the participation of more than 300 research groups worldwide, and subsequently releasing databases for age-related macular degeneration (AMD) and pathologic myopia.
This database was developed under strict quality control, with each image independently annotated by seven specialist physicians and subsequently reviewed by senior experts, garnering significant attention within the industry. In the future, Professor Zhang and his collaborators will release larger-scale databases to foster the robust development of AI in ophthalmology. Beyond academic research, Professor Zhang is also promoting the establishment of consortium and industry standards for AI-enabled ophthalmic databases to facilitate their commercial deployment.
In Professor Zhang Xiulan’s future plans, there are two main areas of focus: first, to continue the in-depth development of AI algorithms for glaucoma, ultimately establishing a multimodal system; and second, to build more high-quality databases, providing premium shared resources for developers and fostering the improvement of China’s ophthalmic AI medical innovation ecosystem.
Currently, top-tier international ophthalmic AI research exhibits two primary characteristics: first, leading medical AI studies leverage massive database resources to train algorithms; second, cutting-edge AI research is exploring algorithm development across different modalities.
Nowadays, AI based on fundus photography has become more mature. In 2017, a large-scale multicenter study led by the Singapore National Eye Centre utilized over 200,000 fundus images to train algorithms capable of screening for diabetic retinopathy, age-related macular degeneration (AMD), and suspected glaucoma, with performance fully reaching human-level accuracy. This represents a highly commendable and instructive successful practice.
Future AI Research in Ophthalmology: Several Frontier DirectionsFirst, the development of advanced, high-performance diagnostic algorithms to create composite systems capable of diagnosing a variety of ocular diseases. Second, the development of algorithms for disease progression prediction, enabling physicians to anticipate future changes in a patient’s condition and facilitate early prevention and treatment. Third, leveraging AI to help physicians identify novel diagnostic approaches by utilizing its powerful computational capabilities to infer correlations between data indicators and diseases.
China has now taken a leading position in AI research, with extensive applied research on AI algorithms. However, to maintain this lead, breakthroughs in foundational and theoretical research as well as original innovation are essential to creating greater value.
Author: Researcher, Baidu AI Industry Research Center
Annotation:
(1) The website for the iChallenge database platform is ichallenge.baidu.com.