Dental Literature Watch · Systematic review · Cariology
AI caries detection: strong numbers, weak evidence, decision support only
TL;DR
- Seventeen systematic reviews, five with diagnostic test accuracy meta-analyses, were pooled in this umbrella review.
- Pooled sensitivity ranged from 0.76 to 0.94 and specificity from 0.85 to 0.91.
- Most systems were deep learning models read on bitewing radiographs and intraoral photographs.
- The evidence is weakened by retrospective datasets, limited external validation, inconsistent reporting, and heavy overlap between the reviews.
- Verdict: not stand-alone diagnostic tools. Use only in decision-support contexts until prospective validation shows an effect on clinical decisions and patient outcomes.
The study
| Element | Detail |
|---|---|
| Design | Umbrella review following PRIOR guidance (PROSPERO CRD420261340728); AMSTAR 2 for quality, corrected covered area for overlap |
| Population | 17 systematic reviews of AI for caries detection and diagnosis, searched to 15 March 2026 |
| Intervention | AI systems, mostly deep learning, applied to bitewing radiographs and intraoral photographs |
| Outcome | Diagnostic accuracy in five bivariate or HSROC meta-analyses; methodological quality and overlap |
| Result | Pooled sensitivity 0.76-0.94, specificity 0.85-0.91; substantial heterogeneity in modalities, lesion thresholds, tasks and metrics |
| Conclusion | High performance under experimental conditions, but not supported as a stand-alone diagnostic tool |
💎 Insight — what this means chairside
Read the sensitivity range as missed lesions
A sensitivity of 0.76 means roughly one lesion in four goes unflagged.
A “clear” AI read on a bitewing is therefore not a reason to skip your own examination.
A sensitivity of 0.76 means roughly one lesion in four goes unflagged.
A “clear” AI read on a bitewing is therefore not a reason to skip your own examination.
Read the specificity range as false alarms
A specificity of 0.85 means up to 15% of sound surfaces can be flagged.
A flag is a prompt to look again, not a restorative decision.
A specificity of 0.85 means up to 15% of sound surfaces can be flagged.
A flag is a prompt to look again, not a restorative decision.
Use it as a second reader
Form your own read first, then check what the software flagged that you did not.
That fits the reviewers' own framing of decision support.
Form your own read first, then check what the software flagged that you did not.
That fits the reviewers' own framing of decision support.
⚠️ The catch
Written from the abstract only
The full text is paywalled, so individual meta-analyses, tools and thresholds are not covered here.
The review's overall message is clear from the abstract, but the detail behind it is not.
The full text is paywalled, so individual meta-analyses, tools and thresholds are not covered here.
The review's overall message is clear from the abstract, but the detail behind it is not.
Good numbers, weak foundations
Most data are retrospective, external validation is limited, and reporting is inconsistent.
Overlap between reviews means the same primary studies are counted repeatedly, which the reviewers say weakens the reliability of the evidence.
Most data are retrospective, external validation is limited, and reporting is inconsistent.
Overlap between reviews means the same primary studies are counted repeatedly, which the reviewers say weakens the reliability of the evidence.
No evidence yet on patient outcomes
The reviewers call for prospective validation showing meaningful impact on clinical decision-making and patient outcomes.
Accuracy in the lab is not the same as better care in the chair.
The reviewers call for prospective validation showing meaningful impact on clinical decision-making and patient outcomes.
Accuracy in the lab is not the same as better care in the chair.
⚡ Bottom line
- Software flags caries on a bitewing → check the flag against your own read; do not restore on the flag alone.
- Software says the film is clear → it does not replace your examination; sensitivity as low as 0.76 means misses.
- Vendor quotes impressive accuracy → ask whether it was validated externally and prospectively on practice-like images.
- Considering AI as a stand-alone reader → not supported by this evidence; decision support only.
🧪 Questions
Tap an answer — instant grade, deciding line, and the trap.
Your practice's imaging software now overlays caries flags on bitewings, and the vendor quotes a sensitivity of 0.94 from retrospective data. Which use matches this umbrella review's conclusion?
The reviewers restrict implementation to decision-support contexts, which means your read first and the software as a second opinion. Screening only the flagged films makes it a stand-alone reader, the use the review says the evidence does not support. Restoring every flag treats a prompt as a diagnosis, and specificity of 0.85-0.91 still leaves false alarms. Photographs only has no support: both bitewings and photographs were the dominant modalities. Trap: the vendor's 0.94 is the top of the range and from retrospective data.
Suppose the true sensitivity of a caries-detection tool were 0.76, the lower end of the pooled range. What follows for a bitewing it marks as clear?
Sensitivity of 0.76 means about 24% of true lesions are missed, so a clear read is not reassurance. Specificity governs false alarms among sound surfaces, not misses, which is why the first option fails. Up to 15% falsely flagged is the correct reading of specificity 0.85 but answers the wrong question for a clear film. Prospective and externally validated is the reverse of the review's finding: retrospective data and limited external validation. Trap: confusing a high specificity with a trustworthy negative.