How AI reads dental X-rays
Most dental AI products on the market today do one job: they look at a radiograph (bitewing, periapical, panoramic or a 3D CBCT scan) and highlight areas a dentist may want to look at more closely. Here is what happens under the hood, in plain language.
1. The software learns from labelled images
These tools use deep learning, a type of machine learning that “learns” statistical patterns from many example images that experts have labelled, for example by outlining caries lesions. After training, the model can make predictions on images it has never seen before [1].
2. It classifies, detects or outlines
Research models do one of three things: classify a whole image or tooth (e.g. “caries present”), detect a finding and draw a box around it, or segment it by outlining the exact pixels. A 2022 systematic review of 42 deep-learning studies on caries detection found all three approaches in use [2].
3. It acts as a second reader – not the diagnosis
Products cleared by the US FDA are typically described as computer-assisted detection (“CADe”) software that aids the dentist. Their official indications state they are not a replacement for the dentist's full review and clinical judgement. You can read this wording in the FDA summaries linked from our tools directory, for example the summary for Overjet Caries Assist [3].
What the evidence says
- Reported accuracy in research studies is promising but varies widely: in the review above, image, object or pixel accuracy ranged between 68% and 99%, and for caries classification on bitewing radiographs it ranged from 87.6% to 95.4% [2].
- Study quality is uneven: only 11 of the 42 studies showed a low risk of bias in all domains, and reporting of performance was heterogeneous, so no meta-analysis was possible [2].
- Experts note that AI has not yet broadly entered routine dental practice, citing limited data, lack of methodological rigour and standards, and open questions on value, ethics and responsibility [1].
What this means for patients and clinicians
AI can be a useful “extra pair of eyes”, especially for consistency in a busy clinic. But a highlighted area is a prompt to look again, not a diagnosis. Ask how the tool was validated, on which X-ray types and age groups, and keep the dentist's clinical examination at the centre of the decision.
Sources
- Schwendicke F, Samek W, Krois J. Artificial Intelligence in Dentistry: Chances and Challenges. J Dent Res. 2020;99(7):769-774. https://doi.org/10.1177/0022034520915714
- Mohammad-Rahimi H, et al. Deep learning for caries detection: A systematic review. J Dent. 2022;122:104115. https://doi.org/10.1016/j.jdent.2022.104115
- US FDA. 510(k) summary K222746 – Overjet Caries Assist. https://www.accessdata.fda.gov/cdrh_docs/pdf22/K222746.pdf