D4Dent Logo

AI in Dental Caries Detection: What the Evidence Says

For dentists

On this page

Introduction

The detection of early carious lesions is one of the routine diagnostic tasks in which conventional methods perform least well. A Cochrane review of 77 studies (15,518 tooth sites or surfaces) on imaging for non-cavitated enamel caries reported a summary sensitivity of 0.47 and specificity of 0.88 across imaging methods; in a hypothetical cohort of 1,000 surfaces with 63% prevalence, this would leave 337 lesions undetected. The certainty of this evidence was rated low [PMID 33720395].

Artificial intelligence (AI), particularly deep learning with convolutional neural networks, has been proposed as a way to close this gap. The question for the practising dentist is not whether such models can achieve high accuracy in a laboratory dataset, but whether they improve diagnosis in the hands of clinicians, and what happens to treatment decisions when they do. This article summarises the published evidence on both questions.

1. How accurate are AI models on bitewing radiographs?

What the evidence shows. Several meta-analyses have now pooled the diagnostic accuracy of AI for caries on bitewings:

  • A 2024 meta-analysis of 21 studies on approximal caries reported a pooled sensitivity of 0.94 and specificity of 0.91. The positive predictive value ranged from 0.15 to 0.87, whereas the negative predictive value ranged from 0.79 to 1.00. The authors concluded that AI is highly reliable in excluding healthy approximal surfaces, but that positive findings should be verified by dental experts to prevent unnecessary treatment [PMID 39396775].
  • A second 2024 review of 14 studies reported a summary sensitivity of 0.87 and specificity of 0.89 (five studies pooled), with sensitivities of 0.84 for dentine caries and 0.71 for enamel caries. Most studies exhibited unclear or high risk of bias, and the authors noted a lack of high-quality, adequately reported and externally validated studies [PMID 38450159].
  • A 2026 meta-analysis of machine learning for proximal caries (28 studies, 11 pooled) reported a pooled sensitivity of 0.84 and specificity of 0.93. Most studies (22 of 28) showed high risk of bias, and the certainty of evidence was low [PMID 41833430].

A broader review of 20 studies across periapical, bitewing and panoramic radiographs reported sensitivities from 0.44 to 0.86 and specificities from 0.85 to 0.98 [PMID 38402191], and an earlier systematic review of 42 deep learning studies concluded that reported accuracy seemed promising while study and reporting quality were low [PMID 35367318].

Compared with dentists. In one development study, a deep neural network tested on 141 bitewings showed an accuracy of 0.80, against a mean of 0.71 for seven dentists, and a sensitivity of 0.75 against 0.36. Its specificity (0.83) was not significantly lower than that of the dentists (0.91). Most dentists showed sensitivities below 0.25 for initial lesions [PMID 32634466]. Three studies in the 2024 bitewing review that compared models with dentists consistently found higher average sensitivity for the models [PMID 38450159].

In our reading, the consistent pattern is that AI adds sensitivity rather than specificity, and its greatest relative advantage is for early enamel lesions, which are exactly the lesions that conventional radiographic assessment most often misses [PMID 33720395][PMID 32634466].

2. Secondary caries and other imaging sources

Around restorations. A model developed on 425 bitewings from seven general dental practices in the Netherlands achieved areas under the FROC curve, a localisation-based accuracy measure, of 0.806 for primary and 0.804 for secondary caries, with F1-scores of 0.689 and 0.719 respectively [PMID 38342368].

On photographs. A systematic review of 19 studies on oral photographs reported F1-scores of 68.3% to 95.4% for professional cameras, 78.8% to 87.6% for intraoral cameras and 42.8% to 80% for smartphone cameras [PMID 37392423]. A 2025 meta-analysis of 45 studies, 33 using radiographs and 12 using clinical images, reported accuracy ranging from 41.5% to 98.6% across 21 AI platforms, a pooled sensitivity of 76% and specificity of 91% (seven studies), with high heterogeneity [PMID 39947783].

In children. A systematic review of AI for early childhood caries reported, for detection studies, accuracy between 78% and 86%, sensitivity from 67% to 96% and specificity from 81% to 99% [PMID 40036840].

An umbrella review of seven systematic reviews concluded that convolutional neural network-based models have considerable potential for accurate, objective caries detection, but that ethical considerations and cautious adoption remain critical [PMID 39206581].

3. Does AI help dentists in practice?

What the evidence shows. The most directly relevant evidence comes from a cluster-randomised cross-over trial in which 22 dentists assessed 20 bitewings each, half with AI support and half without. With AI, the mean area under the ROC curve increased from 0.85 to 0.89, mainly because sensitivity rose from 0.72 to 0.81; specificity was not significantly affected. The increase in sensitivity was found for enamel lesions but not for early or advanced dentine lesions [PMID 34656656].

4. The treatment-decision problem

The same trial found that higher sensitivity came with an increase in non-invasive, but also invasive, treatment decisions [PMID 34656656]. A cost-effectiveness analysis based on the trial data found that AI and no AI showed identical effectiveness (mean tooth retention of 49 years) and nearly identical costs (330 euro in each group), because lesions in the AI group were more often treated invasively. The authors suggested that cost-effectiveness could be improved by supporting not only detection but also subsequent management [PMID 35245626].

An earlier modelling study, which assumed that detected early lesions would be managed non-restoratively, found AI to be more effective and less costly than assessment without AI in most simulations, and stated explicitly that this cost-effectiveness requires dentists to manage detected early lesions non-restoratively [PMID 33198554].

In our reading, this is the central clinical message. An AI system that finds more enamel lesions benefits the patient only if those lesions are managed with non-invasive or micro-invasive care. If additional detections are converted into restorations, the diagnostic gain may be lost.

A practical checklist for AI-assisted caries diagnosis

The following checklist is our interpretation of the evidence above. It does not replace clinical judgement.

  • Use AI output as a second opinion alongside the clinical and radiographic examination, not as the diagnosis [PMID 35367318][PMID 39396775].
  • Verify every positive AI finding clinically and radiographically before planning treatment [PMID 39396775].
  • Record the lesion depth (enamel or dentine) before deciding on treatment; AI support improved sensitivity mainly for enamel lesions [PMID 34656656][PMID 38450159].
  • Consider non-invasive or micro-invasive management for AI-detected enamel lesions, as the economic benefit of AI depends on it [PMID 33198554][PMID 35245626].
  • Treat results from smartphone photographs with particular caution [PMID 37392423].
  • Ask whether the model has been externally validated, since most published studies have not been [PMID 38450159][PMID 41833430].

How D4dent can support dentists

D4dent's AI diagnostics in the doctor workspace is intended to support, not replace, the dentist's own assessment of radiographs and clinical images. In our reading, the evidence summarised above suggests that any such tool is best used as a second reader, with every finding confirmed by the clinician. Dentists using D4dent can also draw on the clinical resources for reference, and on teleconsultation where findings need to be explained to the patient. More about the doctor workspace is available on the D4dent for dentists page.

Frequently asked questions

How accurate is AI at detecting caries on bitewing radiographs?

Meta-analyses report high pooled accuracy. One meta-analysis of 21 studies reported a pooled sensitivity of 0.94 and specificity of 0.91 for approximal caries; another, including 11 studies in its meta-analysis, reported 0.84 and 0.93. However, most included studies were at unclear or high risk of bias, and the certainty of evidence was rated low.

Does AI make dentists better at finding caries?

In a cluster-randomised cross-over trial with 22 dentists, AI support increased the area under the ROC curve from 0.85 to 0.89, mainly by raising sensitivity from 0.72 to 0.81, without a significant change in specificity. The gain was seen for enamel lesions, not for early or advanced dentine lesions.

Can AI lead to overtreatment?

It can. In the same randomised trial, higher sensitivity with AI came with an increase in both non-invasive and invasive treatment decisions. A cost-effectiveness analysis of the trial found that AI and no AI had identical effectiveness and nearly identical costs, because more lesions were treated invasively.

Does AI work on intraoral photographs?

A systematic review of 19 studies found wide variation in performance on oral photographs, with F1-scores of 42.8% to 80% for smartphone cameras, 78.8% to 87.6% for intraoral cameras and 68.3% to 95.4% for professional cameras.

Should AI replace the dentist's diagnosis?

No. The reviews describe AI as an assistant for decisions on the presence or absence of carious lesions, and one meta-analysis advised that positive AI findings should be verified by dental experts to prevent unnecessary treatment.

Limitations to keep in mind

Most diagnostic accuracy studies were at unclear or high risk of bias, and external validation was uncommon [PMID 38450159][PMID 41833430]. Pooled estimates differ between meta-analyses because of differences in included studies, reference standards and lesion depth [PMID 39396775][PMID 38450159][PMID 39947783]. The randomised trial and the cost-effectiveness analyses were conducted in Germany and used one commercial software [PMID 34656656][PMID 35245626][PMID 33198554], so their findings may not apply to other systems or health care settings. None of the cited primary studies was conducted in India. Figures are taken from published abstracts; readers should consult the full texts before citing specific values.

Conclusion

The evidence indicates that AI models can detect caries on bitewing radiographs with high pooled accuracy [PMID 39396775][PMID 38450159][PMID 41833430], and that AI support increases dentists' sensitivity, particularly for enamel lesions [PMID 34656656]. The quality of the underlying studies is, however, generally low, and higher sensitivity has been accompanied by more invasive treatment decisions [PMID 34656656][PMID 35245626]. In our reading, AI is best regarded as a second reader that helps the dentist to find early lesions, and its value depends on managing those lesions preventively.

More in this series

Bring AI diagnostics, patient records, lab orders and clinical resources into one workspace. Create your D4dent doctor account.

References

  1. Walsh T, Macey R, Riley P, Glenny AM, Schwendicke F, Worthington HV, et al. Imaging modalities to inform the detection and diagnosis of early caries. Cochrane Database Syst Rev. 2021 Mar 15;3(3):CD014545. PMID: 33720395
  2. Cantu AG, Gehrung S, Krois J, Chaurasia A, Rossi JG, Gaudin R, et al. Detecting caries lesions of different radiographic extension on bitewings using deep learning. J Dent. 2020 Sep;100:103425. PMID: 32634466
  3. Ammar N, Kühnisch J. Diagnostic performance of artificial intelligence-aided caries detection on bitewing radiographs: a systematic review and meta-analysis. Jpn Dent Sci Rev. 2024 Dec;60:128-136. PMID: 38450159
  4. Carvalho BKG, Nolden EL, Wenning AS, Kiss-Dala S, Agócs G, Róth I, et al. Diagnostic accuracy of artificial intelligence for approximal caries on bitewing radiographs: A systematic review and meta-analysis. J Dent. 2024 Dec;151:105388. PMID: 39396775
  5. Guo K, Yon MJY, Yang Y, McGrath CP, Lam PPY. Diagnostic performance of machine learning-aided proximal caries detection: a systematic review and meta-analysis. J Evid Based Dent Pract. 2026 Mar;26(1):102225. PMID: 41833430
  6. Albano D, Galiano V, Basile M, Di Luca F, Gitto S, Messina C, et al. Artificial intelligence for radiographic imaging detection of caries lesions: a systematic review. BMC Oral Health. 2024 Feb 24;24(1):274. PMID: 38402191
  7. Mohammad-Rahimi H, Motamedian SR, Rohban MH, Krois J, Uribe SE, Mahmoudinia E, et al. Deep learning for caries detection: A systematic review. J Dent. 2022 Jul;122:104115. PMID: 35367318
  8. Abbott LP, Saikia A, Anthonappa RP. Artificial intelligence platforms in dental caries detection: a systematic review and meta-analysis. J Evid Based Dent Pract. 2025 Mar;25(1):102077. PMID: 39947783
  9. Moharrami M, Farmer J, Singhal S, Watson E, Glogauer M, Johnson AEW, et al. Detecting dental caries on oral photographs using artificial intelligence: A systematic review. Oral Dis. 2024 May;30(4):1765-1783. PMID: 37392423
  10. Rokhshad R, Banakar M, Shobeiri P, Zhang P. Artificial Intelligence in Early Childhood Caries Detection and Prediction: A Systematic Review and Meta-Analysis. Pediatr Dent. 2024 Nov 15;46(6):385-394. PMID: 40036840
  11. Chaves ET, Vinayahalingam S, van Nistelrooij N, Xi T, Romero VHD, Flügge T, et al. Detection of caries around restorations on bitewings using deep learning. J Dent. 2024 Apr;143:104886. PMID: 38342368
  12. Mertens S, Krois J, Cantu AG, Arsiwala LT, Schwendicke F. Artificial intelligence for caries detection: Randomized trial. J Dent. 2021 Dec;115:103849. PMID: 34656656
  13. Schwendicke F, Mertens S, Cantu AG, Chaurasia A, Meyer-Lueckel H, Krois J. Cost-effectiveness of AI for caries detection: randomized trial. J Dent. 2022 Apr;119:104080. PMID: 35245626
  14. Schwendicke F, Rossi JG, Göstemeyer G, Elhennawy K, Cantu AG, Gaudin R, et al. Cost-effectiveness of Artificial Intelligence for Proximal Caries Detection. J Dent Res. 2021 Apr;100(4):369-376. PMID: 33198554
  15. Negi S, Mathur A, Tripathy S, Mehta V, Snigdha NT, Adil AH, et al. Artificial Intelligence in Dental Caries Diagnosis and Detection: An Umbrella Review. Clin Exp Dent Res. 2024 Aug;10(4):e70004. PMID: 39206581