Last updated: October 2026
If your dentist has started showing you X-rays with coloured boxes around certain teeth, you have already met dental AI. These systems do not take the picture and they do not make the diagnosis. What they do is read the image in seconds and point to areas that may need a closer look: a shadow between two molars that could be early decay, a drop in bone height around a tooth, a dark spot at the tip of a root.
Dentistry turned out to be a natural fit for this kind of software. Practices take huge numbers of standardized, digital radiographs, and many of the findings that matter are subtle grey-on-grey differences that even experienced clinicians read differently from one another. This article explains how the algorithms work, what they look for, what the published evidence says about their accuracy, and what a highlighted spot on your own X-ray actually means.
Which dental X-rays the algorithms read
Not every dental image is equally suited to AI. Most of the research, and most of the cleared commercial products, focus on two-dimensional intraoral films, because they are taken constantly and in fairly consistent ways.
| Image type | What it shows | Typical AI tasks |
|---|---|---|
| Bitewing | Crowns of upper and lower back teeth and the bone crest between them | Decay between teeth (proximal caries), calculus, bone level, faulty fillings |
| Periapical | One or a few whole teeth, from crown to root tip | Lesions at the root tip, root canal quality, bone loss, caries |
| Panoramic | Both jaws in a single wide image | Tooth numbering and charting, impacted wisdom teeth, larger pathologies, missing teeth |
| CBCT (3D) | Cross-sectional volume of teeth and jaw | Segmentation of teeth and nerves, implant and surgical planning (an active research area) |
The bitewing is the reference radiograph for detecting decay between back teeth, which is why it dominates the caries-detection literature.
From image to highlighted finding: how the software works
The workflow inside a practice usually looks like this. The X-ray is captured by the sensor and saved in the practice imaging software. A copy is sent, locally or to a cloud service, to the AI model. The model returns its findings as an overlay on the original image, often colour-coded by type of condition, within a few seconds. The dentist then reviews the overlay, accepts or rejects each mark, and uses it, if useful, to explain the findings to the patient.
What the model is actually doing
The models are trained by showing them large collections of radiographs in which experienced dentists have outlined each lesion by hand. Over many training rounds, the network learns which patterns of pixels tend to correspond to decay, to healthy enamel, to a metal filling or to bone. Researchers describe three main tasks:
- Classification — does this tooth or surface contain caries, yes or no?
- Object detection — draw a box around each suspected lesion.
- Segmentation — trace the exact pixel outline of the lesion, tooth or bone level, which also allows measurements such as millimetres of bone loss.
The quality of those hand-drawn labels matters as much as the network architecture. If the dentists who labelled the training images disagreed with each other, or if the images came from a single clinic with one brand of sensor, the model can perform worse on X-rays from somewhere else. This is the main reason standards bodies now insist on independent validation data.
What dental AI typically flags on an X-ray
| Finding | Why it is useful | Main caveat |
|---|---|---|
| Caries (decay) | Early proximal lesions are easy to miss by eye | Enamel-only lesions are often managed without drilling; a flag is not a reason to fill |
| Alveolar bone loss | Software can measure bone levels consistently over time | Periodontal diagnosis also needs probing and clinical history |
| Periapical radiolucencies | Dark areas at the root tip can signal infection | Some are healing scars or normal anatomy |
| Calculus (tartar) | Visible deposits below the gumline | Clinical exam remains the main method |
| Restorations & margins | Charts crowns, fillings, implants, root canals; spots open margins | Bright metal can hide decay underneath it |
| Impacted third molars | Quick screening on panoramic images | Surgical decisions need 3D imaging and a specialist |
How accurate is AI at reading dental X-rays?
The most useful numbers come from pooled analyses of independent studies rather than single company announcements. Two 2024 meta-analyses are especially relevant. They looked at different sets of studies and asked slightly different questions, so their results should be read side by side rather than combined.
Caries detection on bitewings overall
A systematic review by Ammar and Kühnisch in the Japanese Dental Science Review screened 935 references and included 14 studies. Only five of them could be pooled for overall caries detection, and the results below come from those five.
AI caries detection on bitewing radiographs: pooled research estimates
95% confidence intervals: overall sensitivity 0.76–0.94; overall specificity 0.75–0.96; dentin sensitivity 0.80–0.87; enamel sensitivity 0.66–0.75. Overall figures pooled from 5 studies. Source: Ammar N, Kühnisch J. Japanese Dental Science Review, 2024; 60:128–136. Sensitivity = share of real lesions found; specificity = share of healthy surfaces correctly left unmarked.
Two points matter for interpretation. The wide confidence intervals show how uncertain the overall estimates still are: the true sensitivity could plausibly be anywhere from the mid-70s to the mid-90s. And performance on decay that has reached the dentin is clearly better than on shallow enamel lesions, which are exactly the ones a second pair of eyes could help with most. The authors also found that while most included studies reported an accuracy of 80% or more, most had an unclear or high risk of bias, and they highlighted a lack of externally validated studies.
Decay between teeth (approximal caries)
A second meta-analysis from Semmelweis University in Budapest, published in the Journal of Dentistry in December 2024, focused only on approximal caries on bitewings. It included 21 of 2,442 screened studies and reported a pooled sensitivity of 0.94 (95% CI 0.78–0.99) and a pooled specificity of 0.91. The very wide lower bound of the sensitivity interval, and the variation between studies, mean this figure should not be read as the typical performance of any one tool. The included studies also did not report enough data to compare AI directly with dentists.
What changed when dentists used AI in a randomized trial
Lab performance is one question; what happens when a real dentist works with the software is another. A cluster-randomized crossover trial from Charité – Universitätsmedizin Berlin, published in the Journal of Dentistry in 2021, had 22 dentists assess bitewings with and without a commercial AI tool. Using AI significantly increased the dentists’ sensitivity, mainly for enamel lesions, while specificity did not change much and accuracy on advanced lesions was essentially unaffected.
The trial also reported a less obvious effect: AI support was associated with more invasive treatment decisions. That matters, because many early enamel lesions can be arrested with fluoride, sealing or infiltration rather than a filling. Finding more lesions is only a benefit if the treatment that follows is proportionate.
Cleared products and the rules around them
In the United States, AI tools that assist with dental image reading are regulated as medical device software and typically reach the market through the FDA’s 510(k) pathway. A clearance means the FDA found the device substantially equivalent to an existing one for a specific intended use. It does not mean the software is infallible, and cleared products do not all rest on the same amount or type of clinical evidence.
| Company / product | Cleared use (as announced) | Type of evidence |
|---|---|---|
| Overjet — Dental Assist | Automatic measurement of bone levels on X-rays (2021); the company has since announced further clearances, including caries detection | FDA 510(k) submissions; performance figures on the company’s website are manufacturer data |
| Pearl — Second Opinion | Detection of caries, calculus and other findings on bitewing and periapical images (2022); a later clearance for caries, periapical radiolucencies and impacted third molars on panoramic images | Standalone performance study and multi-reader, multi-case study, as reported by the company |
| VideaHealth — Videa Caries Assist | Caries detection on radiographs (2022); the later Videa Dental Assist covers a wider range of common findings | Manufacturer-reported FDA clearance study: 43% fewer missed caries lesions and about 15% fewer false detections for dentists using the tool. Results apply to that evaluation, not to AI tools in general |
Regulatory information checked in October 2026. Clearances and intended uses can change; the FDA’s public 510(k) database lists the exact wording for each device.
In the European Union, dental AI software intended for medical purposes may be subject to the Medical Device Regulation (MDR). Depending on its intended purpose and regulatory classification, it may also fall under the high-risk AI provisions of the EU AI Act. Requirements and implementation dates depend on the applicable legal framework. Dental practices should verify a device’s regulatory status and intended use rather than relying on a general claim of certification.
Professional standards are catching up
The American Dental Association has published a small but growing set of documents on the topic: a 2022 white paper (No. 1106) giving an overview of AI uses in dentistry, a 2025 technical report (No. 1109) stressing the need for independent datasets to validate image-analysis algorithms, and ANSI/ADA Standard No. 1110-1:2025, which sets criteria for annotating and collecting 2D radiograph data. The ADA describes the latter as the first U.S. standard on AI in dentistry approved by the American National Standards Institute.
What AI-highlighted cavities mean for patients
A coloured mark on a dental X-ray is a finding to review, not a diagnosis on its own. Its significance depends on where the suspected lesion is, how deep it appears, the condition of the tooth and your overall risk of tooth decay.
Before recommending treatment, a dentist should interpret the image alongside the clinical examination and your dental history. Some early, non-cavitated lesions may be suitable for preventive treatment and monitoring rather than immediate drilling. Others may need a filling, depending on their severity and clinical features. AI can make findings easier to see, but the treatment decision should still rest on an individual clinical assessment.
If your dentist uses AI: questions worth asking
You do not need to understand neural networks to be an informed patient. A few plain questions will tell you how the tool is being used in your care.
- Where exactly is the suspected lesion, and how deep does it appear to be?
- Does the clinical examination support what the software flagged?
- Is treatment needed now, or could preventive care and monitoring be appropriate?
- What are the benefits and risks of each treatment option?
- Is the software cleared or certified as a medical device where you practise, and for which uses?
- Where are my X-rays processed and stored, and is my data used to train the model?
Related reading on Helathora
Limitations of the current evidence
Most studies are retrospective and use relatively small, single-centre image sets. Reference standards differ: some compare the AI with a panel of dentists reading the same X-ray, which is itself imperfect, while very few use histology or clinical follow-up. Staging systems for lesion depth vary between studies, making direct comparisons difficult. Much of the most-cited real-world performance data comes from the companies selling the products. Finally, there is still little evidence on long-term patient outcomes such as tooth survival, overtreatment or cost, although health-economic modelling from the Berlin group suggests AI-supported caries detection may be cost-effective, mainly because fewer lesions go undetected.
The Bottom Line
FAQ
Does AI replace my dentist in reading my X-rays?
No. Cleared dental AI tools are designed as decision support. The dentist reviews the findings, can reject any of them, and remains responsible for the diagnosis and treatment plan.
Will I get more X-rays because my clinic uses AI?
There is no reason it should. The software analyses images that would be taken anyway; how often you need radiographs depends on your age, caries risk and dental history.
Can AI find cavities my dentist would miss?
Sometimes. In a randomized trial, AI support helped dentists find more early enamel lesions. However, many of those lesions can be managed without a filling, so finding them does not automatically mean drilling.
Can the AI be wrong?
Yes. It can mark healthy areas as diseased (false positives) and miss real lesions (false negatives). Image quality, overlapping teeth and metal restorations all affect performance.
Is my dental X-ray data safe?
That depends on the vendor and the clinic. Ask whether images are processed locally or in the cloud, how long they are stored and whether they are used to improve the model. In the EU, health data is protected under the GDPR.
Official Sources
Peer-reviewed research
- Ammar N, Kühnisch J. Diagnostic performance of artificial intelligence-aided caries detection on bitewing radiographs: a systematic review and meta-analysis. Japanese Dental Science Review, 2024; 60:128–136.
- Diagnostic accuracy of artificial intelligence for approximal caries on bitewing radiographs: a systematic review and meta-analysis. Journal of Dentistry, December 2024 (Semmelweis University). Summary in Dental Tribune.
- Mertens S, Krois J, Cantu AG, Arsiwala LT, Pradhan B, Schwendicke F. Artificial intelligence for caries detection: randomized trial. Journal of Dentistry, 2021.
- Mohammad-Rahimi H, et al. Deep learning for caries detection: a systematic review. Journal of Dentistry, 2022.
- Schwendicke F, et al. Cost-effectiveness of artificial intelligence for proximal caries detection. Journal of Dental Research, 2021.
Standards and regulatory information (accessed October 2026)
- American Dental Association. Artificial Intelligence in Dentistry: standards and technical reports.
- U.S. Food and Drug Administration. 510(k) Premarket Notification database.
Manufacturer announcements
- VideaHealth. FDA 510(k) clearance announcement for Videa Caries Assist, May 2022.
- Pearl. FDA clearance for pathology detection on panoramic X-rays (reported by Becker’s Dental + DSO Review).
This article is for general information only and does not constitute medical or dental advice. Diagnostic accuracy figures come from published studies and manufacturer announcements and may not reflect performance in every clinical setting. Always discuss your X-ray findings and treatment options with a qualified dental professional. Helathora is not affiliated with any of the companies mentioned.