A radiologist in a provincial hospital opens a chest CT and, before writing a word, sees a nodule already circled in red. The mark was not drawn by a colleague — it came from a deep-learning model that has been cleared by China's drug regulator to sit inside the diagnostic workflow.
That workflow is no longer a conference demo. By late 2025, the National Medical Products Administration (NMPA, 国家药监局) had granted Class III registration certificates to 207 artificial-intelligence medical devices, and approval volume had topped 40 products a year for three consecutive years, according to figures reported by Xinhua and state-affiliated outlets.
The count matters because a Class III certificate is the strictest category China uses for devices that support, or in some cases help drive, clinical decisions. It is the line between a research prototype and software a hospital can legally load onto a diagnostic workstation.
Key takeaways
- China had approved 207 AI medical-device Class III certificates by December 5, 2025, with 40+ approvals in each of the last three years.
- More than 110 of those are deep-learning-based independent software products cleared by October 2025, spanning CT, X-ray and fundus imaging.
- United Imaging Intelligence (联影智能) holds 20 certificates; Shukun (数坤科技) holds 19; surgical-robot makers such as Edge Medical (精锋医疗) are now inside the count too.
- Regulation is explicit: AI software may assist, but under China's internet-diagnosis rules it cannot replace a physician's signature.
What "Class III" actually buys
China sorts medical devices into three risk classes. Class III is reserved for products that, if they fail, can threaten life — implantable devices, life-support machines, and, increasingly, software whose output feeds a doctor's call. For an AI model, a Class III certificate means it passed a clinical-evidence and quality-system review, not just a bug check. NMPA first published dedicated review guidelines for AI medical devices in 2022 and refined the lifecycle rules again in 2025, explicitly to speed innovation while keeping the safety bar.
Where the software is reading
The cleared products cluster in medical imaging. They triage and flag on CT, X-ray and fundus images — lung nodules, coronary lesions, diabetic retinopathy and similar patterns a tired human eye can miss on a heavy shift. In nuclear medicine, United Imaging Intelligence's chest PET-CT triage software became the first product in that sub-field to win a Class III certificate, and the company says it now holds 20 such certificates overall. The data types NMPA has cleared span at least nine modalities, including CT, X-ray and fundus.
The companies behind the certificates
The 207 is not one vendor's trophy. It is a spread:
- United Imaging Intelligence (联影智能): 20 Class III certificates, including the first nuclear-medicine AI certificate and a multi-disease chest-CT product in the special-review channel.
- Shukun (数坤科技): 19 certificates. Its "Shukun-Kun" (数坤坤) medical large model backs more than 100 scenario apps running from screening through rehab, aimed at standardizing human–machine diagnosis across hospital tiers.
- Edge Medical (精锋医疗): a three-in-one surgical robot platform received its NMPA certificate in March 2026, showing the count now reaches physical surgical systems, not just reading software.
The hospitals already using it
This is not only about certificates on paper. At Beijing Children's Hospital (北京儿童医院), a national "AI pediatrician" (AI儿科医生) works as a research and rare-disease assistant inside multidisciplinary case reviews. At Peking Union Medical College Hospital (北京协和医院), a gait-assessment system for early neurological diagnosis had completed more than 100,000 neuro-function tests by early 2025. Both are examples of the "assist, not decide" pattern the rules are built around.
What the doctor still owns
The vendors are blunt about the boundary. Shukun founder Mao Xinsheng (毛新生) frames the shift as AI moving "from a professional assistive tool to a new-quality productive force in healthcare" — but the company's own chief technology officer, Zheng Chaoming (郑超明), is clearer: AI has no qualification to issue a diagnostic report, and the final sign-off stays with the physician, who carries the liability. China's internet-diagnosis rules already forbid software from impersonating or substituting for a doctor, so the regulatory floor and the vendor posture line up.
Honest limitations
- The headline 207 figure comes from Xinhua and state-affiliated outlets citing NMPA; we did not directly query the primary NMPA device registry to re-count it, so treat it as a late-2025 cumulative figure, not a live total.
- A certificate is not a deployment. The 207 counts approved products, not active hospital installations, and we found no single public tally of how many scans these models actually read per day.
- Per-product accuracy claims vary by disease and dataset; we located no independent, cross-vendor benchmark comparing them head-to-head in real clinics.
- Rules for continuously learning, self-updating models — and for higher-risk treatment-planning AI — are still being written, so part of the pipeline remains ahead of the rulebook.
What readers can do now
- If you operate a clinic, check NMPA's public device registry for the specific product and indication before purchase — certificate number, not a sales deck.
- As a patient, it is reasonable to ask whether an AI assist was used on your scan and, more importantly, who signed the final report.
- Track the 2026 NMPA updates on real-world performance monitoring; "approved" is the start of oversight, not the end.