Caries-detecting AI lifts junior dentists' accuracy. Its weak spot is early decay

Caries-detection software reads a dental X-ray and draws a box around suspected decay, prompting the dentist who might otherwise miss it. In a multicentre reader study published in September 2026, adding one to the workflow raised early-career dentists' tooth-level sensitivity for caries on panoramic radiographs from 67.4% to 82.4%, with essentially no change in specificity — 97.2% to 97.4% [s1]. That is a real gain for less-experienced readers, but independent validation of the same class of tool shows where it falls down: it is far better at confirming a tooth is healthy than at catching early decay [s2].

What the reader study measured

Researchers had twelve early-career dentists — three years or less in practice — and three senior dentists (more than ten years) each read 402 anonymised panoramic radiographs under different conditions: unaided, with AI assistance, and with the AI's standalone output, all against a reference standard set by three experienced dentists using pixel-level annotations [s1]. AI assistance did three things at once. It lifted sensitivity as above; it cut mean interpretation time from 65.12 to 50.37 seconds per image (p=0.003); and it pulled the readers closer together, with inter-reader agreement rising from a kappa of 0.61 to 0.73 [s1]. Case-level sensitivity — whether a patient with any decay was flagged — rose from 84.7% to 93.3% (p<0.001) [s1].

Crucially, the extra catches did not come at the cost of false alarms. Specificity held, which is the result that separates a useful second reader from one that simply makes everyone biopsy more teeth. The standalone system on its own reached a sensitivity of 79.2% (95% CI 76.0–82.4) and an area under the curve of 0.938 (95% CI 0.934–0.941) [s1] — slightly below the assisted human readings, a reminder that the tool worked best as a prompt rather than a replacement.

One number in the paper does not add up and is worth flagging: the reported specificity of the standalone AI, 98.4%, carries a confidence interval (95% CI 76.0–82.4) copied verbatim from the sensitivity line above it, which cannot be correct for a specificity point estimate of 98.4% [s1]. The point estimate is usable; the interval attached to it is a transcription error, not a finding.

The weakness independent testing keeps finding

A separate validation, on bitewing radiographs rather than panoramics, tested the commercial system Diagnocat against two experts on 100 images covering 1,540 tooth surfaces [s2]. Here the asymmetry is stark. Specificity was 94.3% (95% CI 92.4–96.0) and negative predictive value 96.1% (95% CI 94.7–97.3), for an overall accuracy of 91.6% [s2]. But sensitivity was only 73.1% (95% CI 65.9–79.9) and positive predictive value 64.7% — meaning roughly a third of the teeth it flagged did not, on expert review, have decay, and it missed a comparable share of the lesions that were there [s2]. Performance was no better for the early enamel lesions that matter most for prevention: sensitivity 73.3% for enamel caries, with only moderate agreement with the expert consensus (kappa 0.492) [s2].

The authors' reading is blunt: the tool is reliable for ruling caries out, but its lower sensitivity "emphasizes the need for clinician oversight, especially in detecting early-stage disease" [s2]. A scoping review of the field reached a similar place, noting that strong reported accuracy has not yet translated cleanly into everyday clinical decision-making [s3].

Why the two results are not in conflict

They describe the same tool from two angles. In the reader study the AI is a prompt handed to a human who still decides, and a prompt that raises catches without raising false positives is a genuine aid, especially for a junior clinician or a practice without a senior colleague to consult [s1]. Run as an autonomous read on early decay, the same detection style over-calls and under-catches, which is why the independent validation lands on clinician oversight [s2]. It is the recurring shape of clinical AI: a device clearance certifies that software performs a task, not that deploying it improves care, and an accurate reader aid can still fail to change what happens to the patient.

A second, quieter caution runs alongside the benefit. A tool that narrows the gap between readers can also narrow it downward if clinicians defer to it — the deskilling question that colonoscopy AI has already raised. And as with cleared radiology software, the thinness of real-world testing behind marketing claims is why an independent validation that reports the misses as plainly as the catches is the useful kind.

What to watch

Whether these gains hold outside a controlled reading exercise, and whether they change treatment rather than just detection counts. A tooth flagged early is only a benefit if the finding is real and the management differs; an over-call that leads to drilling a sound surface is a harm the accuracy numbers do not capture. The evidence so far supports AI caries detection as a rule-out aid and a leveller for inexperienced readers — not as a substitute for a dentist's eye on the early lesions where it is weakest.

This article is informational and is not medical advice.

Sources

• [s1] "Multicentre Evaluation of AI-Assisted Caries Detection on Panoramic Radiographs Among Early-Career Dentists." International Dental Journal, published online 8 September 2026. https://doi.org/10.1016/j.identj.2026.109812
• [s2] "Evaluating AI diagnostic accuracy in approximal dental caries detection on bitewing radiographs." Clinical Oral Investigations, published online 29 April 2026. https://doi.org/10.1007/s00784-026-06882-z
• [s3] "Assessing the Role of Artificial Intelligence in Caries Detection and Clinical Decision-Making: A Scoping Review." Caries Research, published online 2 February 2026. https://doi.org/10.1159/000550238