AI Caries Detection: Understanding Sensitivity, False Alerts and Dentist Review
Research and guidance checked: 6 October 2026.
Cover photograph: Masked dental professional examining a patient in a dental chair. Illustrative stock photo; not an AI output or a research case. Photo: Ayberk Mirza / Pexels • Pexels licence
A highlighted tooth surface can make a radiograph easier to discuss. It can also draw attention towards a finding that is uncertain. To assess caries AI fairly, separate three questions: can the model identify the study's target, does it help a clinician interpret images, and does it improve decisions for patients?
The 30-second answer
Sensitivity describes how often the study's positive cases are detected.
Specificity describes how often its negative cases are correctly left unflagged.
Neither measure says that every highlighted surface needs treatment.
What the 2026 research reported
A systematic review included 25 eligible studies, with 13 contributing to the meta-analysis. Pooled sensitivity was 0.86, specificity 0.91 and AUC 0.94. Heterogeneity exceeded 90%, indicating major variation across the included evidence. Differences in images, models and reference standards restrict how directly those averages transfer to a local product.
Measure | Reported result | How to read it |
|---|---|---|
Sensitivity | 86% | Detection among reference-positive cases |
Specificity | 91% | Correct negatives among reference-negative cases |
AUC | 0.94 | Discrimination across thresholds; not the chance a flag is correct |
Heterogeneity | I² above 90% | Results varied substantially across studies |
Why an 86% sensitivity can still produce uncertain flags
The following calculation is illustrative. It is not a clinical study, product prediction or claim about the prevalence of caries in your practice. Suppose 1,000 surfaces include 100 reference-positive and 900 reference-negative surfaces. Apply 86% sensitivity and 91% specificity as fixed assumptions.
Illustrative outcome | Surfaces |
|---|---|
Positive surfaces detected | 86 |
Positive surfaces missed | 14 |
Negative surfaces correctly unflagged | 819 |
Negative surfaces incorrectly flagged | 81 |
There would be 167 flagged surfaces, of which 86 are reference-positive: roughly 52%. This demonstrates why prevalence changes the interpretation of alerts. It does not invalidate a useful second reader; it shows why checking each flag matters. The pooled study figures are not necessarily a single operating point that a real device can reproduce.
Dental professional positioning X-ray equipment beside a patient. Illustrative stock photo; not an AI output or a research case. Photo: Gustavo Fring / Pexels • Pexels licence
What happens when a dentist uses the assistance?
In a 2021 randomised reader trial, 22 dentists each assessed 20 bitewing images, split between assisted and unassisted readings. AI assistance increased AUC from 0.85 to 0.89 and sensitivity from 0.72 to 0.81; specificity did not change significantly. Assistance also affected treatment choices, including more invasive proposals. This image-reading study does not establish long-term patient benefit or validate every commercial system.
A useful review sequence
Assess whether the image is suitable for interpretation before examining the overlay.
Review the image yourself, then inspect the system's highlighted areas.
Check disagreements against the clinical information and the system's intended target.
Record the clinician's reasoning when the assistance changes a decision.
Explain uncertainty without presenting the overlay as a verdict.
In a synthetic training example, a clinician sees a faint interproximal area that the software flags. The learning point is to review the area and its context, not to accept a treatment solely because a box appeared. The appropriate decision belongs to the clinical assessment.
Questions worth asking a supplier
Ask whether the test images match your intended bitewing workflow. Request performance by lesion threshold and image quality, and ask how restored surfaces and unusual cases were handled. Find out whether the validation set was independent of training and whether human-plus-AI performance was measured. A model-only score cannot answer every question about a clinician-assisted process.
Frequently asked questions
Are these photographs research images? No. The cover shows a clinical examination and the inline photo shows dental X-ray equipment beside a patient. They illustrate the workflow; neither is a trial image or AI output.
Can a low number of alerts prove the model is good? No. It might indicate an appropriate threshold, a low-prevalence sample or missed findings. Both positive and negative cases need review.
What should a pilot measure? Include missed findings, incorrect alerts, decision changes and the time required to review them. Report the image type and sample size alongside the results.
Sources and image credits
Written for Diamed Engineering. Research findings are attributed above; tables, explanations and labelled synthetic examples are original editorial material.
Photographs are credited beside their use and supplied under the Pexels licence. They do not imply endorsement by the pictured people.
Comments