A new diagnostic accuracy study in Caries Research tested eight AI systems, five commercial and three experimental, on 200 bitewings with 885 restored tooth surfaces. Every system was highly specific but caught only about a third to a half of the secondary caries lesions the reference standard identified, and none performed significantly better than the others on accuracy. For Canadian dental professionals, the finding is a clear signal: a clean AI read around a restoration is not a clean bill of health.
As of October 2026, AI radiograph assistants are increasingly offered to dental practices, and caries next to existing restorations is a challenging call for them. The study, published online on October 9, 2026 by Chaves and colleagues (DOI 10.1159/cre/abuag016), directly compares eight of these tools on that specific task. Here is what it tested, what it found, and what it means for practices in Toronto, the GTA and across Ontario.
Why Secondary Caries Is a Hard Test for AI
Recurrent decay at restoration margins is a familiar diagnostic headache. The authors frame the problem plainly: on a two-dimensional image, true lesions adjacent to restorations are difficult to separate from radiolucencies created by the restoration itself or by normal anatomy. In practice, a radiolucent liner, a cervical burnout zone or an overlapping contact can look a lot like demineralized tooth structure.
The authors note that AI-based clinical decision support systems have been introduced to assist with radiographic interpretation, but also that different AI tools may give different outputs on the same image, and that how they compare against one another had not been clear. This study set out to measure that directly.
How the Study Was Designed
This was a cross-sectional diagnostic accuracy study. The key design elements, as described in the abstract:
- Images: 200 anonymized bitewing radiographs containing 885 restored tooth surfaces.
- Reference standard: a consensus-based classification of each restored surface as non-carious, secondary caries, or indeterminate. Secondary lesions were further staged as enamel-stage or dentin-stage.
- Systems tested: five commercial AI systems (Second Opinion, CranioCatch, Diagnocat, DIO Inteligência and Align X-ray Insights) and three experimental systems built on the Mask R-CNN and Mask DINO architectures.
- Outcomes: sensitivity, specificity, accuracy, positive predictive value and negative predictive value, each with 95% confidence intervals.
- Statistics: comparisons used generalized estimating equations adjusted for clustered data, a method that accounts for related (clustered) observations, such as several surfaces on one bitewing.
What the Researchers Found
The headline result is a consistent pattern across every system, commercial or experimental:
- Specificity was high: 0.957 to 0.986. Sound restored surfaces were rarely flagged as carious.
- Sensitivity was moderate: 0.327 to 0.487. Put simply, for every 10 secondary caries lesions in the reference standard, the systems flagged roughly three to five. The authors note the missed detections were particularly among dentin-stage lesions.
- Accuracy was similar across the board: 0.882 to 0.917, with no statistically significant differences among the systems (p ≥ 0.05). An additional analysis restricted to dentin-stage lesions showed similar patterns.
Where the systems went wrong
The authors traced misclassifications between sound and carious surfaces to three recurring sources: radiographic overlap, radiolucencies related to the restoration, and cervical artifacts. Any clinician who has squinted at a Class II margin will recognize all three.
What the authors concluded
The authors conclude that the algorithms, regardless of architecture or commercial status, showed similar diagnostic capability and a conservative profile that favours specificity over sensitivity. They suggest that more diverse datasets, better reference standards and more explainable models may improve reliability for this task.
What This Means for Ontario and GTA Practices
A conservative tool still needs careful reading. A system with high specificity and moderate sensitivity is more useful for confirming a lesion you already suspect than for clearing a margin you are unsure about.
Pro Tip: Treat an AI "no finding" at a restoration margin as neutral, not reassuring. If your clinical exam, tactile findings or the patient's history point to recurrent decay, let that drive the decision, and record your own radiographic interpretation in the chart rather than relying on the software's output alone.
Three practical implications follow from the study's numbers:
- Brand choice may matter less than workflow. No system was significantly more accurate than the others in this study. Before paying a premium for one platform over another on detection claims, ask how it fits your imaging software, your charting and your team's review habits.
- Dentin-stage misses deserve attention. The authors highlight that missed detections were particularly among dentin-stage lesions, which typically are the ones more likely to need intervention. That makes human review of restored surfaces more important, not less.
- Use the flags as a second look, not a first filter. Because sound surfaces were rarely flagged, an AI flag on a restored surface is a good prompt to re-examine that site carefully.
Pro Tip: If you are trialling an AI radiograph tool, pull 20 to 30 recent bitewings with restorations where you later confirmed or ruled out recurrent caries clinically, and see how the tool reads them. Your own cases are the most relevant test set you have.
Canadian regulatory context
The study abstract does not address Canadian availability or Health Canada licensing for any of the systems tested. Health Canada's guidance on Software as a Medical Device (SaMD) treats software as a medical device when, among other criteria, it is intended for a medical purpose, and lists diagnosing, screening or detecting a disease or condition among those purposes. Confirm a product's Health Canada licensing status with the vendor before purchasing. Whatever tool you use, the radiographic interpretation in your chart is your clinical judgment; the Royal College of Dental Surgeons of Ontario (RCDSO) is the place to check professional expectations for your practice.
Limits of what we know
The full text of this paper is behind a subscription, and this summary is based on the published abstract. The results apply to the 200 bitewings and the specific system versions tested; AI products may be updated after testing, and performance on your own imaging hardware and patient population may differ. The study measured detection against a consensus-based reference standard, which is itself an imperfect benchmark for a difficult diagnosis.
EBIKO Dental will continue monitoring research on AI-assisted diagnosis and what it means for Canadian practices. For more dental industry news and supply resources, visit EBIKO Dental.
Sources
- Chaves ET, Domburg JT, Romero VHD, et al. AI-Based Clinical Decision Support Systems for Secondary Caries Detection and Staging on Bitewings: A Multi-Algorithm Comparison. Caries Research, published online October 9, 2026. doi:10.1159/cre/abuag016 (Europe PMC record)
- Health Canada. Guidance Document: Software as a Medical Device (SaMD): Definition and Classification
Frequently Asked Questions
Q: How accurate are AI tools at finding secondary caries on bitewings?
In a 2026 Caries Research study of eight AI systems on 885 restored surfaces, accuracy ranged from 0.882 to 0.917, with no significant differences among systems. Specificity was high (0.957 to 0.986) but sensitivity was moderate (0.327 to 0.487), meaning many lesions were missed.
Q: Which AI system performed best for recurrent caries detection?
None stood out on accuracy. The study found no statistically significant differences in accuracy among the five commercial systems (Second Opinion, CranioCatch, Diagnocat, DIO Inteligência and Align X-ray Insights) and three experimental systems tested.
Q: Can Ontario dentists rely on an AI "no caries" result around a restoration?
Not on its own. With sensitivity below 0.5 in this study, a negative AI read should be weighed alongside your clinical exam and your own radiographic interpretation, and you should confirm any tool's Health Canada licensing status before using it.