Dental AI X-Ray Reports: How Often Is a Flagged Problem Real?

A dentist can now hand you a printed report in which software has marked every tooth on your X-ray. Diagnocat, one of the companies selling these tools, tells patients its report is "a reliable second opinion powered by advanced AI technology."¹ Its own blog describes clinics running international patients' scans through the software before they travel, then sending back a report with "potential treatment needs" and suggested treatment order.²

So the report can end up shaping a treatment plan before anyone has looked in your mouth. That makes one question worth answering with published numbers rather than marketing: when dental AI flags a problem on your X-ray, how often is the problem real?

What a US clearance actually covers

Each US Food and Drug Administration (FDA) clearance reviewed here names specific findings on one image type. Each is for use alongside a dentist, who reads the X-ray with the software's marks on screen. None is cleared to diagnose on its own.

Overjet's caries tool is the clearest example. In the company's FDA test, it found 72.0% of decayed surfaces on 352 bitewing X-rays.³ It wrongly flagged 1.9% of healthy surfaces, a specificity of 98.1%. When 13 US dentists read the same images with and without it, their sensitivity rose from 57.9% to 76.2%. Their specificity slipped from 99.3% to 98.4%.

A 98% figure sounds close to perfect, but it is counted per tooth surface, and Overjet's test set held about 20 surfaces per X-ray. At that rate, a set of healthy teeth would collect roughly one false flag for every two or three bitewing images. That figure is this publication's arithmetic from the FDA summary, not a number Overjet publishes.

Pearl reports its tools differently. Its periapical caries tool found 88% of lesions in a test of 1,250 images.⁴ When 21 dentists used it, the summary reports a "slight increase" of 0.16 false marks per image. Pearl's panoramic tool, cleared in November 2025, found 77% of decay on panoramic films.⁵

Diagnocat's US clearance is narrower than its report

The FDA's database lists two clearances under Diagnocat's US company, DGNCT LLC.⁷ One, from September 2025, is a viewer for 3D scans. Its summary says it makes no diagnostic claims.⁸ The other, from January 2026, is for one finding: dark areas at the tips of roots on cone-beam CT, a 3D dental scan.⁶ On that task it scored 85.4% sensitivity and 99.1% specificity.

That is a much smaller scope than the report patients see. Diagnocat's patient page promises "a clear description of each tooth, warning signs" and dentist comments.¹ The same page adds that features "may not be available or approved for use in all regions."

By comparison, the same FDA database lists 13 clearances under Overjet's name.⁹ It lists six detection clearances for Pearl, including the panoramic tool.¹⁰ So a patient abroad may receive a full-mouth Diagnocat report, of which the US has reviewed one finding on one scan type. The rules in the clinic's own country are a separate question this article did not check.

Independent tests of Diagnocat disagree

Panoramic X-rays are the wide films that show the whole mouth in one image. Diagnocat's dental-tourism blog lists them among the scans it analyses.² Three published studies tested Diagnocat on them, and their results point in different directions.

The 2022 Warsaw study is small, and its authors declared no conflict of interest.¹¹ It found Diagnocat cautious about decay: the software missed more than half of it, but almost never flagged a healthy tooth. It also missed 61% of root-tip lesions.

The 2025 study, published in Applied Sciences, found nearly the opposite pattern.¹³ Diagnocat caught 99.2% of conditions the experts marked. But its overall specificity was 8.7%, which the authors read as "a tendency to overdiagnose." For decay, specificity was 47%. Its agreement with the experts on decay scored −0.15 on a scale where zero means chance-level agreement.

Both figures are published, peer-reviewed and specific. They cannot both describe the same tool in the same conditions. The studies used different image sets, different reference panels and different years of the software, and none of them explains the gap. This article does not pick between them.

The 2023 study is the most favourable. It reported "almost perfect agreement" on most findings, but low sensitivity for decay and root-tip lesions.¹² Eight of its 12 authors list Diagnocat Inc. as their employer. Its lead author is a scientific adviser to the company, according to the paper's own disclosure.

Precision: the number that answers your question

Accuracy studies usually report two numbers. Sensitivity is the share of real problems the software finds, and specificity is the share of healthy teeth it correctly leaves alone. Neither answers the question a patient holding a report has. That question is precision: of the spots the software flagged, how many were real?

Precision depends on how common the problem is. When most teeth are healthy, even a small false-flag rate can put many wrong marks next to a few real ones.

A 2024 study of 600 panoramic X-rays shows the effect, using a different product from the ones above.¹⁴ It tested Carestream Dental's CS Imaging software against two human readers. For root-tip lesions, the software scored 99.2% specificity. Yet only 72.5% of its root-tip flags matched the readers, so roughly one in four did not.

Most of the Diagnocat studies above report sensitivity and specificity, not precision. A patient usually cannot look up the figure that matters most for the tool their clinic uses.

Better detection can also mean more drilling

Even a tool that finds decay accurately can change what happens next. A randomised trial at Charité hospital in Berlin tested this directly.¹⁵ Twenty-two dentists read bitewing X-rays with and without an AI tool called dentalXrai Pro.

With AI help, their sensitivity rose from 0.72 to 0.81, while their specificity did not change significantly. The gain came from early decay in the outer enamel, not from deeper lesions.

Both non-invasive and invasive treatment decisions rose. The authors concluded that AI "may also increase invasive treatment decisions." The software maker was founded by two of the trial's authors, Falk Schwendicke and Joachim Krois.¹⁶ Their trial still published the downside.

The case for the tools

The evidence does not say dental AI is unreliable. A 2025 umbrella review in PLoS One pooled 20 studies of AI decay detection.¹⁷ It found pooled sensitivity of 85% and specificity of 90%.

In the US clearance studies above, dentists found more disease with the software than without it.

The same review also sets limits on its own numbers. Only 20 of 137 studies it found reported enough data to pool. The decay rate in the test sets was 27%, which the authors call lower than in real communities. They found signs of publication bias.

At that 27% rate, the review calculates that a positive AI result means a 79% chance of real decay.¹⁷ Put the other way round, about one positive flag in five would be wrong.

What the numbers leave out

Who decided the right answer. Every accuracy figure above measures the software against human readers. The Warsaw and 2025 studies used different panels and got opposite results. None of the studies here checked a flag against what was found when the tooth was opened.

The image you actually had. A clearance for bitewings says nothing about a panoramic film. Diagnocat's detection clearance in the US covers 3D scans, not panoramic films.

Who reads the report. The US clearances assume a dentist reviews each mark. A report emailed to a patient, or used to build a quote abroad, may skip that step. None of the sources here measures how often that happens.

Your clinic's version and settings. Diagnocat assigns each finding a likelihood score from 0 to 100%.¹³ Where a clinic sets the cut-off changes how many flags appear. Patients are not told the setting.

How to use an AI report on your X-ray

Ask which software produced the report, and which image type it read. Then check whether that tool is cleared for that finding on that image. In the US, the FDA's 510(k) database is public and searchable by company name.

Ask the dentist to point to each flagged spot and say whether they agree with it, because a flag they do not stand behind should not appear on your quote.

For any flag of early decay, ask whether it can be watched or treated without drilling. That is the decision the Berlin trial showed AI can push the wrong way.

Before crowns, veneers or root canals, get the same X-ray read by a dentist who is not selling the treatment. An AI report can be one of the inputs to that second look. It cannot be the second look.