An ophthalmologist squints at a stack of cross-section images of a patient's retina, looking for the faint swelling that signals the start of irreversible vision loss. The scan holds more slices than any human can check in a busy clinic. Now a software tool reads the same stack in seconds and points to the exact layer where something looks wrong.
That tool is MIAS-3000, an AI diagnostic system for ophthalmic OCT (optical coherence tomography, 光学相干断层扫描) images, built by a team at the University of Electronic Science and Technology of China (UESTC, 电子科技大学). In late October 2025 it became the first OCT-based eye AI to clear China's strictest medical-device bar.
A scanner that sees what tired eyes miss
OCT is the ultrasound of the eye — it builds a three-dimensional picture of the retina and the choroid beneath it, the layers where age-related macular degeneration (AMD) and diabetic eye disease do their damage. Most approved eye AI in China works on flat colour photographs. OCT adds depth, which means more information and a far harder computational problem.
MIAS-3000 was developed at UESTC's School of Medicine under Professor Chen Xinjian (陈新建). On 27 October 2025 it received a Class III medical-device registration certificate from China's National Medical Products Administration (NMPA, 国家药监局), registration number 20253212107. UESTC describes it as the world's first Class III-certified AI system for reading medical OCT scans — a category that, until now, had no approved product at that regulatory level anywhere.
What the software actually does
The system takes the OCT slices and, using a deep-learning model, flags the lesions that matter for AMD: drusen, fluid buildup, and abnormal new vessels. It is aimed at adults aged 50 and over, and it returns an assisted read on whether mid-stage or more advanced macular degeneration is present.
Crucially, the team paired the software with the hardware. By embedding the model inside automated OCT devices, they lowered the skill needed to operate the scanner and read the result. The goal is to let county hospitals, border-region clinics, optometry shops, and health-check centres run precise retinal screening without shipping every image to a specialist in a big city.
The numbers behind the clearance
Regulators do not hand out Class III certificates lightly, so the trial figures are the part worth reading closely. In clinical testing reported by UESTC:
- Sensitivity reached 95.5% (95% CI 92.5%–97.5%) — the system caught the large majority of true cases.
- Specificity reached 92.1% (95% CI 89.6%–94.2%) — it rarely cried wolf on healthy eyes.
- Diagnostic agreement with expert readers came in at 94.7% (95% CI 93.5%–95.7%).
- Sub-group checks showed 98.0% sensitivity on the wet, neovascular form of AMD, the faster-blindness version that most needs catching early.
The same study reported that the software shortened reading time for clinicians. For a country where more than 700 million people have some form of eye disease and where retinal conditions are now a leading cause of blindness, a faster, consistent first read is the whole point.
Why OCT, and why now
China's eye-care system has the wrong shape for its population: too few specialist ophthalmologists, concentrated in coastal cities, serving an ageing country. AMD and diabetic retinopathy build silently; by the time a patient notices blurred vision, damage is often permanent. The cheapest lever is earlier, wider screening — and that is exactly what a model embedded in a routine scanner enables.
This is the deployment logic behind a certificate, not a demo. A Class III clearance means the product can be used in real clinical decision-making, not just as a research curiosity. UESTC frames the wider aim as a full-cycle eye-health loop covering screening, diagnosis, treatment, and follow-up.
MIAS-3000 also sits inside a wider Chinese wave of cleared diagnostic AI. Vendors such as United Imaging Intelligence (联影智能) have stacked more than a dozen Class III certificates across CT, MRI, and X-ray reading. The common thread is not a single breakthrough model but a regulatory pipeline that now lets hospitals buy and bill for AI reads. That is what turns research code into a line item in a clinic's budget, and it is the quiet reason China's medical-AI deployments keep outpacing the demo stage.
According to deployment reading from the site owner's observatory, China's upload of clinical-grade AI into primary care is the trend to watch: the value is less in a flashy model and more in putting a consistent second reader into every understaffed clinic.
Honest limitations
The sensitivity and specificity figures are vendor- and trial-reported through UESTC's own announcement; this article did not locate a separate peer-reviewed publication of the pivotal trial, and the "world's first" claim is the university's characterization rather than an independent ranking. The clearance covers an assisted read for AMD in patients 50 and over, not a general-purpose eye diagnosis. Real-world accuracy outside the trial cohort, and the cost of rolling the embedded devices into grassroots clinics, are not yet published.
What readers can do now
- If you are over 50 or have diabetes, ask your eye clinic whether they offer OCT retinal screening — the test exists even where the AI does not.
- Clinic builders should treat Class III clearance as the minimum bar; a model without it is a research tool, not a clinical one.
- Track how many grassroots Chinese clinics actually deploy embedded OCT AI over the next year — deployment, not certification, is the number that will matter.