New AI Reads Your X-Ray Reports Without Sending Data Away

When a radiologist reads your chest X-ray, they type up a report in plain sentences — something like "no evidence of pneumonia" or "possible small nodule." Hospitals want computers to read thousands of those reports at once, so they can spot patterns, check quality, or find patients for research studies. The problem is that today's best AI tools for this job are giant models accessed over the internet, and hospitals are understandably reluctant to send patient text to an outside company.

A research team led by Kai Yu built a much smaller alternative. Their model, ChestPheNoT, comes in sizes between 0.5 and 3 billion parameters — tiny by modern standards, small enough to run on hardware a hospital already owns. For every report, it outputs three things: the finding (say, a lung nodule), a status (present, absent, or uncertain), and the exact words from the report that support that call. That last part is what makes it auditable — a doctor can check the receipts instead of trusting a black box.

The team trained it using labels generated by an older system plus a much larger 72-billion-parameter model, then refined it with extra practice, the same way you'd train a new employee on a mentor's notes. Across three sets of human-checked reports, the 3B version was a bit weaker than its teacher on familiar data but held up better when tested on reports from a different hospital. It beat the older industry standard, CheXbert, by 2 points on that harder test, and more than 99% of the quoted evidence it produced could actually be found in the original report.

The honest limits: the model is roughly half a billion to 3 billion parameters — capable, but not as sharp as the giant models on familiar material. It only handles chest imaging language, and it is still a research paper, not a product your hospital can switch on today. Still, it signals a shift: private, local AI that shows its work is becoming competitive with the cloud.