There is a category distinction buried in every regulatory clearance for dental imaging software, and it explains most of what practices find confusing about the current market.
Assistive tools are cleared. Autonomous ones are not.
What the distinction means
An assistive tool produces an output a clinician reviews. The clinician remains the diagnostician; the software is a second set of eyes with no authority. Clearance for these is comparatively routine, and dozens of products have obtained it.
An autonomous tool produces a diagnosis. Nothing in dental imaging has been cleared to do that, and there is no visible near-term path.
The practical consequence is that every product on the market — regardless of how it is marketed — is legally a second opinion offered to a dentist who remains fully responsible for the read. Vendors know this. Their documentation says so plainly. Their marketing frequently implies something else, and the gap between the two is where practices get into trouble.
The measurement problem nobody has solved
Evaluating these tools is harder than it sounds, because the ground truth is contested.
To say a model found a lesion a dentist missed, you need to know the lesion was there. The usual reference standard is a panel of experienced clinicians, which means the model is being measured against exactly the kind of judgment it is supposed to improve on. Where a tooth is subsequently opened, you have better evidence — but only for the teeth someone decided to open, which is not a random sample.
Dr. Julia Kowalski notes that this is not a dental problem specifically. “Every diagnostic model in medicine has this issue,” she says. “Dentistry is unusual in how rarely it is mentioned in the sales conversation.”
The thing that has actually changed my practice is that something notices when a lesion is bigger than it was eighteen months ago.
Where the tools are clearly good
Two areas, consistently.
Detection of interproximal caries on bitewings is genuinely strong, and the disagreements in the literature are about calibration thresholds rather than capability.
And comparison across time — flagging change between a current radiograph and a prior series — is a task software does well and humans do badly, mostly because humans do not reliably have the prior image open.
Dr. Nathan Whitfield, an oral surgeon, describes the second as the underrated one. “The detection headlines get the attention,” he says. “The thing that has actually changed my practice is that something notices when a lesion is bigger than it was eighteen months ago.”
Where they are not
Anything requiring clinical context the image does not contain. Symptom history, thermal testing, mobility, occlusal findings — none of it is in the radiograph, and a model confident about a periapical finding without it is confident about something narrower than a diagnosis.
That is the whole reason for the assistive designation, and it is a reasonable place for regulators to have drawn the line.
