 |
| Title |
Fully Automatic AI Tracing of 2D Lateral Cephalograms Can Aid Orthodontic Assessment, But Current Evidence Does Not Support Replacing Expert Verification |
| Clinical Question |
In patients who receive orthodontic treatment with 2D lateral cephalograms, how accurate is fully automated AI software identifying cephalometric landmarks and reproducing cephalometric measurements when compared to expert reference tracings? |
| Clinical Bottom Line |
Fully automated AI tracing of 2D lateral cephalograms can assist orthodontic assessment, but current evidence does not support replacing expert verification. Hendrickx et al. reported an average landmark error of approximately 1.4 mm, below the commonly used 2-mm threshold, while Rauniyar et al. found that 77% of landmark predictions were within 2 mm of the reference location. Coronel-Zubiate et al. found generally small or inconclusive differences between AI and manual measurements, with a small statistically significant difference for ANB but not SNA; these findings do not establish clinical equivalence. Bias, heterogeneity, and inconsistent reference standards limit confidence in the evidence, supporting AI use with clinician review and correction. |
| Best Evidence |
(you may view more info by clicking on the PubMed ID link) |
| PubMed ID |
Author / Year |
Patient Group |
Study type
(level of evidence) |
| #1) 38895901 | Hendrickx, 2024 | 34 studies (27 lateral cephalogram; 7 CBCT) | Meta-Analysis | | Key results | This study included 20,451 2D image entries, summed from Table 2; reused/training datasets prevent a unique-patient total.
For automated landmark identification on 2D cephalograms from 34 studies, there was a mean radial error of 1.43 mm (95% CI 0.95–1.91; I² = 0%), which is below the clinically acceptable 2mm threshold. AI automated landmark identification was time-efficient (<1 min), but accuracy varied by landmark. No pooled sensitivity, specificity, PPV, NPV, or likelihood ratios were reported. For 3D images, no meta-analysis was performed, but the mean error of landmark detection ranged from 1.0 to 5.8 mm. There was no direct comparison with semi-automated tracing. | | #2) 37496553 | Rauniyar, 2023 | 7 studies; 4,184 image/sample entries including training data; 5 studies pooled | Meta-Analysis | | Key results | Unique tested-patient total and demographics for this dataset are unclear.
Five studies were pooled, producing a successful detection rate of 77% within 2 mm of the reference landmark location (95% CI: 69%–85%). Approximately 23% of predictions therefore fell outside that tolerance. The 2-mm threshold defines a successful individual landmark prediction; the review did not establish a minimum overall percentage that qualifies a system as clinically acceptable. Thus, 77% should not automatically be interpreted as sufficient for unsupervised clinical use. Heterogeneity was substantial (I² = 94.47%; reported p = 0.00), indicating considerable variation among studies. This p-value concerns heterogeneity, not superiority over manual tracing. The pooled outcome was landmark detection success, not sensitivity; pooled specificity, predictive values, and likelihood ratios were not reported. Semi-automated methods were not directly compared. | | #3) 42698514 | Coronel-Zubiate, 2026 | 22 studies (2020–2025) | Meta-Analysis | | Key results | This study included mainly single-centre orthodontic samples using lateral cephalograms or CBCT. Unique patient total not reported in the main text.
Across 16 cephalometric measurements, most pooled differences between AI and manual tracing were small or inconclusive. SNA was not statistically significant (Hedges’ g = 0.11; 95% CI: −0.03 to 0.24; p = 0.14). ANB showed a small statistically significant difference (g = 0.28; 95% CI: 0.03–0.53; p = 0.03), but its clinical importance remained uncertain. These standardized differences do not establish greater accuracy or interchangeability. Evidence certainty ranged from low to moderate. Sensitivity and specificity were not pooled. | |
| Evidence Search |
PubMed:
cephalometr* AND "artificial intelligence" AND (review OR meta-analysis) |
Comments on
The Evidence |
Hendrickx et al.: two reviewers screened the studies, the search was broad, and the authors used QUADAS-2 to assess bias. However, many of the included studies had some form of bias (selection), and some datasets were reused, so we are less confident in the results.
Rauniyar et al. had a lot of heterogeneity between studies and only included English-language papers. It did not provide explicit information about whether the review was done independently by more than one person. In addition, successful cases were calculated using the total samples rather than clearly separated test sets. All of this weakens the pooled estimate.
Coronel-Zubiate et al. used duplicate assessment, QUADAS-2, and GRADE, which strengthens its methods. Even so, the certainty of the evidence was only moderate. The reference standards were inconsistent, and some measurements showed substantial heterogeneity. The authors also tested 16 outcomes without adjusting for multiple comparisons, so some significant results could be due to chance.
Finally, the three reviews draw on overlapping primary studies, so they shouldn't be treated as independent replications of each other. It's also worth remembering that a lack of statistical significance doesn't prove two methods are equivalent.
|
| Applicability |
The findings apply best to standard orthodontic radiographs utilized in the studies (cephalograms). We can conclude that AI can expedite tracing, but a clinician still needs to check the landmark placements and the measurements/interpretation. However, it is important to note that it's unclear how well these tools work on atypical facial or skull structures, or on images from different imaging systems. Using this technology in a practice also means having the software, training staff to use it, and allowing time to correct its mistakes. Finally, these reviews don't show that AI is cost-effective or that it leads to better outcomes for patients. It is also unclear the considerations with HIPAA when uploading patient images into AI systems. |
| Specialty/Discipline |
(Oral Medicine/Pathology/Radiology) (Orthodontics) |
| Keywords |
Artificial intelligence, lateral cephalometry, landmark identification, automatic tracing, and manual tracing
|
| ID# |
3770 |
| Date of submission: |
10/15/2026 |
| E-mail |
patels36@livemail.uthscsa.edu |
| Author |
Sneh Patel |
| Co-author(s) |
Taylor Reusche, Ricardo Guerrero |
| Co-author(s) e-mail |
reusche@livemail.uthscsa.edu, guerrero5@livemail.uthscsa.edu |
| Faculty mentor/Co-author |
N/A |
| Faculty mentor/Co-author e-mail |
N/A |
Basic Science Rationale
(Mechanisms that may account for and/or explain the clinical question, i.e. is the answer to the clinical question consistent with basic biological, physical and/or behavioral science principles, laws and research?) |
post a rationale |
| None available | |
|
Comments and Evidence-Based Updates on the CAT
(FOR PRACTICING DENTISTS', FACULTY, RESIDENTS and/or STUDENTS COMMENTS ON PUBLISHED CATs) |
post a comment |
| None available | |
|
|
