A 68-year-old man in a county clinic slides his chin onto a small camera. Forty seconds later, a screen shows his retina has been read for signs of diabetes damage — with no ophthalmologist in the building.
Across town, a shopper at an optical chain gets the same one-minute check before picking out a pair of glasses.
Neither encounter looks like the hospital visit that diabetic-eye screening used to require.
The camera that reads the back of the eye
The retina is the only place in the body where blood vessels and nerves can be seen directly without cutting anything open. A photograph of the back of the eye, a fundus image, therefore carries traces of conditions far beyond vision: diabetes, hypertension, some forms of stroke risk.
Diabetic retinopathy (糖尿病视网膜病变) is the leading cause of blindness in working-age adults. In China, with roughly 114 million people living with diabetes, the disease affects somewhere between a quarter and more than a third of them — tens of millions of individuals at risk of sight loss. The bottleneck was never the camera. It was the eye doctor. Public reporting has long noted that China has only about 1,100 physicians specifically trained to read retinal diagnostic images, mostly clustered in big-city tertiary hospitals, while the population that needs screening is everywhere.
A first certificate, then a stack
Airdoc (鹰瞳科技) was founded in Beijing in September 2015 and lists on the Hong Kong Stock Exchange under code 02251.HK, having gone public on November 5, 2021 as the first Hong Kong-listed medical-AI company.
Its core product, Airdoc-AIFUNDUS, carries a milestone that matters more than the branding: in August 2020 it became the first artificial-intelligence ophthalmic auxiliary-diagnosis software in China to receive a Class III medical-device certificate from the National Medical Products Administration (NMPA, 国家药监局). Class III is the strictest category, reserved for devices that carry the highest risk and therefore the highest evidence bar. The certificate covered auxiliary diagnosis of diabetic retinopathy.
The pipeline has since widened:
- AIFUNDUS 1.0 — diabetic retinopathy, NMPA Class III, August 2020.
- AIFUNDUS 2.0 — multi-disease fundus, adding conditions such as retinal vein occlusion; NMPA Class III approval in April 2025.
- AIFUNDUS 3.0 — aimed at pathological myopia and retinal detachment, with research in a late stage but its registration application paused in May 2025 because, in the company's own filing, market awareness and acceptance were still low.
Separately, standalone software for glaucoma (June 2020) and cataract (January 2022) received Class II certificates from the Shanghai drug regulator. According to Airdoc's 2026 interim report, the Retina-AI system now supports the identification of 55 types of lesions and disease risks.
How it actually gets used
The product is not parked in research hospitals. Airdoc states its retinal-AI tools run in grade-A hospitals, community clinics (社区诊所), health-checkup centers, insurance companies, optometry centers (视光中心) and pharmacies (药房), and says it has served tens of millions of users.
The commercial footprint is visible in the company's own 2022 annual results, which two Chinese financial outlets reported in detail: revenue of about 114 million yuan (roughly $16 million), customer count rising from 244 to 397, an average of 2,371 service points per month, and 4.33 million retinal screenings completed that year.
Underneath the screening products sits the Wanyu (万语) medical large model (大模型), described by the company as built on more than 30 million real clinical records and over 800 evidence-based knowledge graphs. Airdoc also cites a real-world study published in The Lancet Digital Health with the Zhongshan Ophthalmic Center (中山大学中山眼科中心) covering 14 retinal abnormalities at expert-level accuracy.
Why this matters for deployment
The interesting question is not whether an algorithm can read a photo. It is whether the algorithm changes who gets read. Airdoc's chief medical officer, Chen Yuzhong (陈羽中), has framed the tool publicly as a way to add capacity to stretched grassroots clinics rather than to replace eye doctors — the camera extends a tertiary-hospital capability to a county or a pharmacy that would never have run a screening program otherwise.
That is the real "AI落地" story here: the shift from a specialist-led exam to a frontline, low-staffed touchpoint. It only works because the device is a regulated medical product with a certificate number, not a consumer app guessing at your health.
Honest limitations
- Auxiliary, not autonomous. The NMPA certificate is for auxiliary diagnosis (辅助诊断). A physician still makes the call, and the system's value is bounded by image quality — a blurry or poorly framed shot degrades the result.
- Geographic concentration. Deployments skew toward eastern cities, checkup chains and insurance-wellness channels. True rural community-clinic penetration remains limited, which is exactly where the doctor shortage is worst.
- Uneven rollout. The much-discussed multi-disease expansion (2.0) only cleared NMPA in 2025, and the 3.0 plan was paused. Headlines about "55 risks" describe capability, not yet routine clinical use for all of them.
- Payment gap. Screening is often paid out of pocket, via employer wellness budgets or insurers, rather than covered by public reimbursement, which limits reach for lower-income patients.
- Data sensitivity. Health data flowing through pharmacies and insurers raises consent and security questions that vary by operator.
What readers can do now
- If you have diabetes, or are over 40, ask your community clinic or a major optical chain whether a retinal-AI check is available — one photo is fast and non-invasive.
- Treat a "normal" result as a reason to stay alert, not to skip the eye doctor. If you have symptoms such as blurred vision or floaters, see a specialist; the tool assists, it does not replace.
- For employers or insurers evaluating vendors: request the NMPA certificate number and class before adopting any retinal-AI service, and confirm which conditions it is actually cleared to flag.