In brief: recent systematic and umbrella reviews report promising performance for AI-based caries detection, while also highlighting heterogeneity, retrospective datasets and limited external validation. One headline accuracy number is therefore not enough for procurement or clinical use.

What does a caries model actually detect?

Within AI in dentistry, caries is one of the most studied image tasks. Models may use bitewings, periapical radiographs, intraoral photographs or smartphone images. They can perform binary detection, localization or segmentation, but they do not automatically know lesion activity, patient risk or the most appropriate treatment.

What does the evidence show?

A 2026 umbrella review synthesizing 17 systematic reviews found generally promising diagnostic performance but substantial heterogeneity in imaging modalities, lesion thresholds, tasks and reporting. Repeated reliance on retrospective datasets and limited external validation weakened confidence in broad generalization. A 2025 meta-analysis also found strong performance while emphasizing variability between studies.

Bitewing versus intraoral images

Bitewing radiographs are central to proximal-caries research. Intraoral and smartphone photography can support screening and community use, but lighting, angle and lesion stage matter. A 2025 systematic review of smartphone AI found encouraging feasibility but variable sensitivity for early or non-cavitated lesions.

Detection is not risk prediction

Detection asks whether a lesion signal is visible now. Risk prediction estimates future likelihood and may combine history, behavior and clinical factors. The second task requires a different dataset, outcome definition and validation strategy.

How to evaluate a product

  1. Define intended use: detection, staging or treatment recommendation.
  2. Match imaging modality and devices to your environment.
  3. Ask for external validation rather than internal benchmark only.
  4. Review results by lesion threshold and image quality.
  5. Check uncertainty handling and clinician override.
Important boundary

Finding a suspicious region is not the same as deciding to restore a tooth.

Treatment requires lesion activity, risk, examination and patient context.

For the wider imaging context, see AI in dental radiology. For operational follow-up after a clinician confirms a finding, see AI in dental practice management.

Selected evidence

  1. AI for dental caries detection: umbrella review, 2026.
  2. Accuracy of AI in caries detection: systematic review and meta-analysis, 2025.
  3. AI for binary dental caries diagnosis, 2026.
  4. AI-driven smartphone imaging for caries detection, 2025.

Educational content only; it does not replace clinical diagnosis or treatment planning.