AI cephalometric landmarking: inside 2 mm, most of the time

Automatic landmarking is the most studied AI task in orthodontics. A 2021 meta-analysis found deep-learning tools placed about 80% of landmarks within 2 mm. Two FDA-cleared AI cephalometric software products report passing their own bench thresholds. The studies behind both carry real caveats.

Editorial · not sponsored L3 · Peer-reviewed study

Evidence used: Systematic review and meta-analysis (abstract), two FDA 510(k) summaries
Checked: One PubMed abstract, two 510(k) summary PDFs. Oct 31, 2026 (ET)

Short answer · editorial read

Mostly, as a first draft. Pooled accuracy sits near the 2 mm threshold that studies commonly use, but the underlying studies carry a high risk of bias, and every tracing still needs clinician review.

Strongest source we used

1 Vendor claim
2 Regulatory record
3 Peer-reviewed study
4 Randomized trial

Cropped lateral cephalometric radiograph showing the jaws and teeth, tinted in warm tones. — Lateral cephalometric radiograph by ANUG, Wikimedia Commons, CC BY-SA 4.0; cropped and tinted. Illustration only; not an AI output.

01 · Claims

What vendors and sponsors claim

Their own words or their own testing. Treated as claims until independently checked.

K231396: 99% of landmarks met the endpoint. The sponsor's summary reports a study of 21 landmarks against three orthodontic specialists, counting up to 2.0 mm as a pass. It says "99% of landmarks were identified, meeting the study endpoint", against an acceptance criterion of at least 85%. Claim · sponsor summary
K243005: all acceptance criteria met. The sponsor's summary reports passing mean radial error (MRE) criteria of at most 1.5 mm on lateral and 2.5 mm on frontal (PA) cephalograms, with automatically detected landmarks reviewed by orthodontists. Claim · sponsor summary

02 · What we checked

What we independently checked

What we could confirm from public records and peer-reviewed abstracts or full texts in this pass. Not a clinical study.

Pooled accuracy is around the 2 mm line. Schwendicke et al. (2021) pooled 19 studies (2017–2020), all using convolutional neural networks, mostly on 2D lateral cephalograms. The proportion of landmarks detected within 2 mm was 0.799 (95% CI 0.770–0.824). Checked · PubMed abstract
The evidence base is biased toward public datasets. Most of the pooled studies used publicly available datasets. The authors rate risk of bias as high, with applicability concerns for most studies, mainly in data selection and how the reference test was set. Checked · PubMed abstract
Both clearances are real FDA records. K231396 was decided Jan 31, 2024 and K243005 on May 30, 2025, both Substantially Equivalent under regulation 892.2050. K231396 is indicated for patients 14 and older. Checked · 510(k) summaries
Neither clearance needed clinical testing. The K243005 summary states no clinical testing was required. The K231396 summary describes a bench comparison against specialists. Both describe outputs as dependent on clinician review. Checked · 510(k) summaries

03 · Limits

Limits and what we could not verify

Gaps, caveats and anything we did not read.

Different metrics, not comparable. K231396 reports a share of landmarks within 2.0 mm; K243005 reports mean radial error; the meta-analysis pools both kinds of measure across different datasets. These figures cannot be ranked against each other. Limit · metrics
Test sets not described publicly. Neither summary gives the number of images, where they came from, or the image equipment used in the public text we read. Not verifiable · test data
Landmarks are not diagnoses. None of these sources tests whether AI tracing changes treatment decisions or outcomes. Gap · outcomes
Review evidence is from 2017–2020. Newer models may perform differently. We did not read the full text or any later review in this pass. Unchecked · newer studies

Buyer takeaway: ask for per-landmark error on images from your own machine, especially for landmarks that are hard to place by hand, and keep a quick check step for every tracing before measurements are used.

04 · Sources

Sources

Deep learning for cephalometric landmark detection: systematic review and meta-analysis (Schwendicke, Chaurasia, Arsiwala, et al.) Clinical Oral Investigations 2021;25(7):4299-4309 · PubMed abstract · Jul 2021 Partly read Abstract read. Full text not read in this pass.
510(k) Summary K231396 (PDF) Sponsor-written summary, filed with FDA · Issued Jan 28, 2024 Read Oct 31, 2026 Sponsor-written summary; performance section read.
510(k) Summary K243005 (PDF) Sponsor-written summary, filed with FDA · 2025 Read Oct 31, 2026 Sponsor-written summary; accuracy study section read.