In a county hospital in Jiangxi, one ophthalmologist reads dozens of fundus photos a day. An AI checks each in about four seconds and flags the ones a tired eye might miss. This is not a conference demo. The system cleared China's strictest medical-device bar and is installed in more than a hundred hospitals.
For anyone watching AI in medicine, the useful question is not "can a model see a lesion?" It is "did it survive the regulator?"
The product that entered the clinic
Tencent Miying (腾讯觅影) launched in August 2017 as Tencent's first medical-imaging AI. Unlike many research projects, it pursued — and won — China's Class III (三类) medical-device registrations, the highest regulatory tier, where a product may be used as an auxiliary diagnostic system in hospitals.
Three certificates anchor it:
- Chronic glaucoma-like optic neuropathy fundus image auxiliary diagnostic software (注册证号 国械注准20223211140) — screens for glaucoma from fundus photos.
- Colorectal polyp electronic endoscope image auxiliary detection software (国械注准20233210707) — flags polyps during colonoscopy; listed on 1 June 2023 and also recognised as an innovative device.
- Pneumonia CT image auxiliary triage and assessment software (国械注准20213210612) — approved August 2021.
Tencent says it holds six medical-device registrations across respiratory, eye and digestive diseases, making it one of the first internet companies with multi-disease, multi-modal medical AI certificates.
What the certificates actually buy
A Class III (三类) certificate is the difference between "interesting tool" and "something a hospital can procure." The clinical figures Tencent reports, drawn from product requirements and multi-centre trials, are specific:
- Glaucoma AI: sensitivity 96%, specificity 92%; cuts missed-diagnosis rate by about 20% and misdiagnosis by about 2%; returns a result in roughly four seconds.
- Colorectal AI: raises polyp detection by more than 20% and lowers missed-polyp rate by more than 20%.
By early 2025 the glaucoma module alone had reached over 100 medical institutions, per Southern Metropolis Daily — a real, scaled deployment rather than a pilot.
The COVID proving ground
Miying's pneumonia AI earned its credibility under pressure. In February–April 2020, at Zhongnan Hospital of Wuhan University (武汉大学中南医院), the system helped read lung CTs for more than 24,000 patients across two months; at makeshift Leishenshan (雷神山) Hospital, reports landed within about 30 minutes of a scan. That emergency use became the basis for the 2021 Class III certificate.
Where it helps most
The clearest value is in primary and rural care, where specialists are thin. In Guangdong and Jiangxi, Miying powers regional eye-health screenings; one Jiangxi county hospital reported finding 13 glaucoma cases in a single month that a thinly staffed clinic might have missed. Deployment is hybrid: tier-3 hospitals run it locally for speed and data control, while grassroots sites use the cloud version to lower the barrier.
What it cannot replace
Miying is an auxiliary (辅助) system — a doctor still signs the diagnosis. Its reported accuracy figures are vendor numbers from trials, not independent head-to-head studies. And its coverage is limited to the diseases it was certified for; other modules remain research-use only.
What readers can do now
- If you are a clinician in China, treat a Class III (三类) certificate as the real procurement bar — model size matters less than regulatory clearance.
- If you build health AI, the moat is the registration pathway and clinical evidence, not the architecture.
- If you follow policy, watch how China's drug regulator (NMPA) reviews AI devices — the rules are still hardening.
Honest limitations
Core facts (Miying launch August 2017; the three Class III registration numbers and dates; six total registrations; glaucoma sensitivity/specificity and missed-diagnosis figures; colorectal detection uplift; 24,000+ patients in Feb–Apr 2020 at Zhongnan Hospital; 100+ institutions by early 2025) come from tencentmiying.com, cloud.tencent.com, GMW (光明网) and Southern Metropolis Daily / nfnews. Registration numbers are published by Tencent's official materials. All clinical performance figures are Tencent's vendor-reported numbers from product requirements and multi-centre trials and have not been independently audited. No RMB amounts are stated. Analysis is current to 28 September 2026.