How good is AI at spotting cavities? A review of 25 studies gives a number

Software that examines a dental X-ray and flags possible decay has moved from research demo to commercial product over the past few years. The claims made for it vary considerably. A systematic review published in BMC Oral Health in February 2026 set out to establish what the pooled evidence actually shows.

What was reviewed

The team screened the literature on AI systems performing binary caries diagnosis — the yes-or-no question of whether decay is present — using either intraoral photographs or dental radiographs. Twenty-five studies met the inclusion criteria, and thirteen carried enough comparable data to be pooled statistically.

The numbers

• Sensitivity: 0.86 (95% CI 0.82–0.89) — of teeth that genuinely had decay, the AI identified 86%
• Specificity: 0.91 (95% CI 0.88–0.94) — of sound teeth, it correctly cleared 91%
• AUC: 0.94 (95% CI 0.92–0.96) — a measure of overall discrimination, where 1.0 is perfect

A useful split emerged between the two input types. Models working from intraoral photographs were better at finding decay — higher sensitivity at 0.88. Models working from radiographs were better at ruling it out — higher specificity at 0.92.

What the authors were careful to say

The review notes “substantial heterogeneity and limitations in study quality and reference standards” and recommends cautious interpretation.

That caveat deserves unpacking, because it is the most important sentence in the paper. Studies differed in how they defined ground truth — what counted as a real cavity against which the AI was scored. Many were conducted on curated image sets rather than in live clinics. Performance on a clean research dataset routinely exceeds performance on the messier images real practices produce.

The authors’ conclusion is that AI should serve as a complementary tool rather than a replacement for clinical judgement.

Why 86% is both good and not enough

Turn the sensitivity figure around: roughly one in seven genuine cavities was missed. As a second pair of eyes on an image a dentist has already examined, that is useful — it catches things fatigue or a rushed schedule might let through. As the only reader, it is not adequate.

The specificity number matters too. A system that flags healthy teeth as decayed leads to unnecessary treatment, which for a patient is worse than a missed early lesion that would have been caught at the next check-up.

What this means for patients

If your clinic uses AI-assisted imaging, it is reasonable to ask how it is used — whether it flags areas for the dentist to examine, or whether findings are acted on directly. The first is the appropriate use as the evidence currently stands.

The examination that matters is still the one performed by a dentist, with imaging and software supporting the diagnosis rather than making it.