In orthodontics, the best-studied AI task is marking landmarks on side-view skull X-rays so that jaw and tooth measurements can be taken automatically. A 2021 systematic review of 19 studies pooled the share of landmarks placed within 2 millimeters of an expert's mark at 79.9 percent, and rated the evidence as consistent but at high risk of bias. Reviews of AI for extraction decisions and treatment planning describe accuracy that often exceeds 80 percent against expert choices, but nearly all of those studies are retrospective and come from single centers.
Key takeaways
- The 2021 meta-analysis of cephalometric landmark detection found a pooled 79.9 percent of landmarks within 2 mm, with very high heterogeneity between studies.
- A public benchmark built from three medical centers saw top challenge teams reach a mean detection rate near 75.7 percent, lower than many single-study figures.
- A 2025 meta-analysis of AI for extraction decisions pooled seven studies and 6,261 patients, with sensitivity of 70 percent and specificity of 90 percent.
- A 2026 review of 19 planning studies concluded the systems should act as adjunctive aids, not autonomous planners.
Where AI enters orthodontic work
Orthodontics deals with the alignment of teeth and jaws. Three kinds of tasks attract most of the research. The first is cephalometric analysis, in which a clinician marks anatomical landmarks on a lateral skull radiograph and calculates angles and distances from them. The second is the decision about whether a treatment plan should include removing teeth. The third is broader planning, such as classifying the skeletal pattern or choosing a treatment approach. AI models are trained on past cases labeled by experts, so they learn to approximate expert judgment. They do not independently establish what is best for a patient.
What the landmark-detection evidence shows
A systematic review and meta-analysis in Clinical Oral Investigations by Schwendicke and colleagues covered 19 studies published from 2017 to 2020. All used convolutional neural networks, 15 on two-dimensional images and four on three-dimensional images. The pooled proportion of landmarks within 2 mm was 0.799 (95% confidence interval 0.770 to 0.824). The two-dimensional subgroup reached 0.792 and the three-dimensional subgroup 0.870. Heterogeneity was very high (I-squared of 99 percent), meaning the studies disagreed widely.
The authors rated risk of bias as high for most studies, mainly in how data were selected (16 studies) and how the reference marks were produced (18 studies). Most studies relied on the same public dataset, which limits how far the numbers generalize. Their conclusion was that accuracy is relatively high and consistent, and that robustness and generalizability still need to be demonstrated.
A harder test set
A 2024 preprint on arXiv, titled "Is the Problem Solved?", describes the CL-Detection dataset: 600 lateral radiographs with 38 landmarks each, collected from three medical centers using different equipment. It reports the results of the top ten teams in a 2023 challenge. The best methods reached a mean detection rate of about 75.7 percent and a mean radial error of about 1.52 mm. The authors concluded that the problem is not fully solved, though highly accurate automatic localization looks plausible. The paper is a preprint, so it has not necessarily gone through full peer review, but the contrast is instructive: results on data from several scanners and hospitals are lower than the figures many single-dataset studies report.
Extraction decisions and treatment planning
A 2025 meta-analysis in BMC Oral Health by Ziaei and colleagues, published October 9, 2025, examined AI models that predict whether orthodontic treatment needs tooth extraction. It pooled seven studies from six countries covering 6,261 patients and reported sensitivity of 70 percent and specificity of 90 percent, with high heterogeneity and funnel plots suggesting possible publication bias. In plain terms, the models missed about three in every ten cases that experts judged to need extractions, and wrongly flagged about one in ten of the cases that did not.
A systematic review by Baxmann and colleagues, published online June 12, 2026 in BMC Oral Health, covered 19 studies of AI and rule-based systems in diagnosis and planning, searched through October 15, 2025. Accuracy against expert decisions often exceeded 80 percent, and in some studies 90 percent. Risk of bias was mostly moderate to serious because of retrospective designs, single-center samples, limited external validation, and the use of expert opinion as the reference standard. The authors recommended treating the systems as adjunctive aids and called for multicenter studies with external validation before routine use.
The evidence at a glance
| Task | Headline result | What limits it |
|---|---|---|
| Landmark detection on X-rays | 79.9 percent of landmarks within 2 mm (19 studies) | High risk of bias, one shared dataset, very high heterogeneity |
| Multi-center benchmark | About 75.7 percent detection rate, 1.52 mm mean error | Preprint; tests only landmark placement |
| Extraction decisions | Sensitivity 70 percent, specificity 90 percent (seven studies) | High heterogeneity, possible publication bias |
| Diagnosis and planning | Often above 80 percent agreement with experts (19 studies) | Retrospective, single-center, expert opinion as the reference |
Agreement with an expert is a different measure from a better outcome for the patient. None of the headline figures above measures whether a person ended up with a healthier bite, shorter treatment, or fewer complications.
Questions to ask about an AI-assisted plan
- Which part of the plan, measurement or decision, came from software?
- Did the orthodontist review and adjust the software's output?
- Was the tool tested on patients and equipment like the ones in this clinic?
- How were patient images stored and shared with the software provider?
This article gives general information about published research. It is not medical or dental advice, and it does not recommend any treatment, product, or provider. Decisions about orthodontic care belong with a patient and a licensed clinician.
FAQ
What is cephalometric analysis?
It is a set of measurements taken from a side-view radiograph of the head. A clinician marks landmarks on bone and soft tissue and calculates angles and distances from them, and AI tools try to place those marks automatically.
How accurate is AI at placing landmarks?
A 2021 meta-analysis pooled 79.9 percent of landmarks within 2 mm of expert marks. A 2024 multi-center benchmark reported a mean detection rate of about 75.7 percent, so accuracy depends on the data.
Can AI decide whether teeth should be removed?
Research models predict extraction decisions with moderate sensitivity (70 percent pooled) and high specificity (90 percent). Reviewers describe such systems as aids to a clinician, not a replacement for one.
Does the research show better treatment results with AI?
The reviews cited here measure agreement with experts or landmark error, not patient outcomes. They call for prospective, multicenter studies with external validation.
Related reading
- AI and dental education
- AI note-taking trials in clinics
- AI and the future of work
- Deloitte AI Institute
More in AI in health.
Last verified: October 2026
Sources
- Schwendicke F et al., Deep learning for cephalometric landmark detection: systematic review and meta-analysis, Clinical Oral Investigations, 2021
- Zhang H et al., Deep Learning Techniques for Automatic Lateral X-ray Cephalometric Landmark Detection: Is the Problem Solved? arXiv preprint 2409.15834, September 2024
- Ziaei S et al., Accuracy of artificial intelligence in orthodontic extraction treatment planning: a systematic review and meta analysis, BMC Oral Health, October 9, 2025
- Baxmann M, Zsoldos M, Kárpáti K, AI in orthodontic treatment planning: a systematic review comparing learning approaches, BMC Oral Health, June 12, 2026