In-house artificial intelligence-assisted cephalometric landmark tracing improves senior dental students’ accuracy: a counterbalanced crossover study

Introduction

Cephalometric analysis remains a fundamental component of orthodontic diagnosis, treatment planning, and assessment of craniofacial relationships. By characterising skeletal relationships, dental inclinations, craniofacial morphology, and growth-related patterns, cephalometric tracing continues to inform both clinical decision-making and orthodontic education. Despite the widespread adoption of digital tracing software, accurate landmark localisation remains a demanding task. Manual tracing is time-consuming and prone to inter- and intra-examiner variability, with outcomes influenced by operator experience, radiographic quality, anatomical superimposition, and the clarity of landmark definitions (1).

Artificial intelligence (AI), particularly deep learning (DL), has increasingly been applied to automate and standardise cephalometric landmark identification. Convolutional neural networks (CNNs) and related architectures have demonstrated performance comparable to experienced orthodontists in selected landmark detection tasks (2,3). Landmark localisation is commonly assessed using mean radial error (MRE) and success detection rate (SDR). Across recent studies, MRE values for AI-assisted cephalometric landmarking commonly fall within approximately 1.1–1.8 mm, while SDRs are frequently reported at around 70–87% within 2 mm and above 90% within 4 mm (1,2,4). These performance ranges have led to increasing interest in AI-assisted cephalometry as a means of improving efficiency, consistency, and reproducibility in orthodontic workflows (5).

Different CNN-based approaches have been proposed to improve landmark detection, including networks using residual connections, attention mechanisms, and multiscale feature extraction (3). However, performance is not uniform across all landmarks. Well-defined skeletal landmarks are generally detected more reliably, whereas selected dental, soft-tissue, or low-contrast landmarks remain more challenging because of anatomical variability, overlapping structures, and reduced radiographic definition (1,4). This variation is particularly important in educational settings, where students may struggle not only with identifying the correct anatomical point, but also with translating conceptual landmark definitions into accurate radiographic placement.

Commercial cephalometric platforms such as CephX, Dolphin Imaging, AudaxCeph, and WebCeph have incorporated AI-based landmark detection into clinical and educational workflows (5). One comparative study of manual, semi-automated, and fully automated cephalometric tracing reported intraclass correlation coefficient (ICC) values above 0.90 for the evaluated orthodontic indices, while Bland-Altman analyses showed satisfactory agreement between the automated methods and manual tracing (6). Nevertheless, available evidence does not indicate that one commercial platform is consistently superior across all landmarks, measurements, and clinical situations. Performance appears to vary according to the specific landmark and cephalometric variable, and selective orthodontist review remains advisable for challenging landmarks and treatment-critical measurements (2,3,7,8). Current AI-assisted cephalometric systems are therefore better understood within the American Dental Association’s concept of augmented intelligence (AuI): AI used to assist clinicians while retaining human intelligence within the task and enhancing rather than replacing professional judgement (9).

The potential value of AI-assisted cephalometry extends beyond technical accuracy to include workflow efficiency, reproducibility, and educational support. Manual cephalometric tracing may require 10–20 minutes per radiograph, whereas AI-based systems can generate landmark predictions within seconds (2,4). By reducing examiner-dependent variation related to fatigue, experience, and tracing conventions, AI may support more reproducible analysis, particularly for novice clinicians and students (7,10,11). In orthodontic education, this raises an important possibility: AI-generated landmark suggestions may function as an educational scaffold, helping learners compare their own judgement with a consistent starting point while still requiring them to inspect, accept, or correct the final landmark position.

Although AI-assisted cephalometry has gained increasing attention, several methodological and educational gaps remain (5). Many models are trained on retrospective single-centre datasets, external validation is inconsistent, and system performance depends heavily on training data quality, annotation consistency, population diversity, malocclusion range, and imaging device characteristics (4,12,13). In addition, much of the existing literature has focused on AI-only technical performance or comparisons against expert tracings (14,15). Educational evidence remains comparatively limited, particularly regarding how AI assistance affects learner performance when used as an adjunct to landmark tracing rather than as a fully automated replacement (11).

Another limitation of many available AI-assisted tracing workflows is their dependence on commercial or externally developed systems (5). Although such platforms may be clinically useful, they often provide limited transparency regarding model development, training data, preprocessing, and deployment. In-house development of AI-assisted tools may offer an alternative educational model by allowing trainees and educators to engage directly with data preparation, model training, validation, and implementation. This may strengthen AI literacy while producing locally relevant tools that can be evaluated within real teaching workflows.

The present study evaluated an in-house AI-assisted cephalometric tracing system developed and deployed locally to support landmark identification by senior undergraduate dental students. Using a counterbalanced crossover design, students traced the same lateral cephalograms under two conditions: unaided manual tracing and AI-assisted tracing. Student-generated coordinates were compared with a three-specialist reference standard. The primary aim was to determine whether AI-assisted tracing improves landmark localisation accuracy compared with unaided manual tracing. A secondary aim was to assess whether the effect of AI assistance was consistent across tracing sequence, thereby addressing potential order and task-familiarisation effects. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0104/rc).

Methods

Study design

This study was designed as a counterbalanced crossover comparative validation study evaluating whether AI-assisted cephalometric landmark tracing improves senior undergraduate dental students’ landmark localisation accuracy compared with unaided manual tracing. Student-generated tracings were compared against a three-specialist reference standard. The study included two tracing conditions: unaided manual tracing and AI-assisted tracing.

The 24 students were divided into two equal sequence groups. Twelve students completed unaided manual tracing first, followed by AI-assisted tracing. The remaining 12 students completed AI-assisted tracing first, followed by unaided manual tracing. This counterbalanced structure was used to balance the order in which the two tracing conditions were completed.

In the unaided manual condition, students placed cephalometric landmarks independently. In the AI-assisted condition, students traced the same landmark set with access to AI-generated landmark suggestions, which they could accept or adjust before saving their final tracing.

An overview of the model-development and student-facing evaluation workflow is presented in Figure 1.

Figure 1 Overview of model development and student-facing evaluation. AI, artificial intelligence; MRE, mean radial error; SDR, success detection rate.

Participants

The study included 24 senior undergraduate dental students in the final clinical stage of the dental programme. Students were recruited from the undergraduate dental programme using convenience sampling. Students were eligible if they had received prior teaching in orthodontic cephalometric landmark identification. No additional participant-level measures of previous orthodontic experience or academic performance were collected. Recruitment and data collection started on 1 February 2026 and concluded at the end of May 2026. Students were included in the final analysis only if they completed both tracing conditions and had complete coordinate data for all included images and landmarks. The study was conducted in an educational dental setting, and all tracing data were collected as part of the study-specific cephalometric tracing exercise.

Image sample

The final image set consisted of 11 de-identified lateral cephalograms, all of which were used for the specialist reference tracings and both student tracing conditions. Images were included when all required landmarks were visible, an identical filename or identifier and the complete set of eight anatomical landmarks were present across the specialist reference, unaided manual, and AI-assisted datasets, and scale calibration was available and consistent. Images were excluded before the final analysis if these cross-dataset matching requirements were not met or calibration was unavailable or inconsistent. Selection was therefore based on landmark visibility, reliable matching across datasets, and calibration completeness; the images were not purposively stratified by malocclusion pattern or formally graded according to anatomical complexity.

Landmark set

Eight anatomical landmarks were included in the accuracy analysis. These were Sella (S), Nasion (N), A point (A), B point (B), upper incisor tip (UIT), upper incisor apex (UIA), lower incisor tip (LIT), and lower incisor apex (LIA). These were recorded as S, N, A, B, UIT, UIA, LIT, and LIA.

The eight landmarks were predefined to provide representation of both skeletal and dental structures, comprising four skeletal landmarks (S, N, A, and B) and four dental landmarks describing upper and lower incisor position (UIT, UIA, LIT, and LIA). They were consistently available across the specialist reference, unaided manual, and AI-assisted datasets and encompassed landmarks with differing localisation difficulty. Other cephalometric landmarks were outside the predefined model output and student tracing protocol and were therefore not included.

Two artificial calibration points, the upper scale-bar endpoint (SCALEU) and lower scale-bar endpoint (SCALEL), represented the upper and lower ends, respectively, of the 10-mm scale bar. The pixel distance between these endpoints was used to derive the image-specific millimetre-per-pixel conversion factor. These points were used only for calibration and were not included as anatomical landmarks in the accuracy analysis.

Development of the in-house AI-assisted cephalometric tracing system

An in-house AI-assisted cephalometric tracing system was developed to generate landmark suggestions for the AI-assisted tracing condition. The system was programmed, trained, and deployed locally using Python and open-source libraries. It was designed to identify the same eight anatomical landmarks used in the student comparison: S, N, A, B, UIT, UIA, LIT, and LIA. Two additional scale endpoints were included during model development to support image-specific calibration, giving a total model output of ten points. The final configuration used 1,024×1,024 image preprocessing, a 10 mm scale reference, ten output channels, and Gaussian heatmaps with a sigma value of 4.0.

The model was implemented as a heatmap-regression network based on a U-Net architecture with a 34-layer residual network (ResNet-34) encoder. During training, each cephalogram and its landmark coordinates were converted into target heatmaps, and the network was trained to predict one heatmap per landmark. The training workflow included an automated annotation audit to verify that all expected landmarks were present, that landmark names matched the model label list, and that duplicate landmark entries were absent. The dataset was split reproducibly into training, validation, and testing subsets using a fixed random seed. The final configuration used 399 images for training, 50 for validation, and 50 for testing. The 11 cephalograms used in the student-facing evaluation were entirely separate from the 499 images used for model training, validation, and internal testing and were not used at any stage of model development. All training, validation, and testing images originated from the same local dataset; no independent external dataset was used for model validation.

Model training was performed using PyTorch. The network was optimised using Adam with decoupled weight decay (AdamW) with a learning rate of 0.0003 and weight decay of 0.01. Training was run for up to 100 epochs with early stopping patience of 20 epochs. Because the batch size was set to 1, gradient accumulation was used over eight steps to stabilise optimisation. The training script saved the best-performing checkpoint and stored the landmark-channel order in the checkpoint to prevent mismatch between training and inference.

For inference, the trained model generated one heatmap for each landmark. Landmark coordinates were obtained from the predicted heatmaps using peak detection with optional sub-pixel refinement based on a local weighted centroid around the heatmap maximum. Predicted coordinates were then mapped back to the original image coordinate space. Model performance during development was assessed using MRE and SDRs within 2 mm and 4 mm, calculated overall and per landmark.

The trained model was incorporated into a local graphical application that allowed a user to upload a lateral cephalogram, generate predicted landmarks, view the annotated image, calculate selected cephalometric measurements, and export landmark coordinates and measurements. As shown in Figure 2, the application included a landing interface for image upload and analysis, followed by an output screen displaying the traced cephalogram, calculated values, and downloadable results.

Figure 2 In-house AI-assisted cephalometric tracing application. Top: landing screen showing the image upload interface and analysis function for lateral cephalograms. Bottom: example output screen showing the AI-traced cephalogram with automatically generated landmark points, displayed cephalometric measurements, and options to download the landmark coordinates as a CSV file and the annotated image in PNG format. ANB, A point-Nasion-B point angle; CSV, comma-separated values; NB, Nasion-B point line; PNG, Portable Network Graphics; SN, Sella-Nasion line; SNA, Sella-Nasion-A point angle; SNB, Sella-Nasion-B point angle.

To distinguish the unmodified model output from students’ final AI-assisted tracings, a post hoc descriptive analysis was conducted using the raw AI-generated coordinates. This analysis included one deterministic prediction for each of the 88 image-landmark combinations represented by the 11 cephalograms and eight anatomical landmarks. Raw AI coordinates were compared with the same three-specialist reference standard and converted to millimetres using the same image-specific calibration factors. Because the model generated only one coordinate set per image, raw AI predictions were not duplicated across the 24 students, and no inferential comparison with the participant tracing conditions was performed.

The software environment used Python with NumPy, Pandas, OpenCV, Matplotlib, SciPy, PyTorch, TorchVision, tqdm, PyYAML, and Streamlit.

Reference standard

Three orthodontic specialists independently traced the included cephalograms. Each specialist was blinded to the other specialists’ tracings and completed landmark placement without access to AI-generated suggestions or other automated landmark-localisation aids. Their files contained the same image identifiers and landmark labels. Specialist data were checked for completeness, duplicate entries, and matching image-landmark combinations.

For each included image and anatomical landmark, the x and y coordinates from the three specialists were averaged to create the three-specialist reference standard. Inter-specialist agreement underlying the reference standard was evaluated for each anatomical landmark after application of the same image-specific calibration factors. Variability was summarised as the Euclidean distance between each specialist coordinate and the corresponding three-specialist mean coordinate and was reported as the mean, standard deviation (SD), and maximum. Because the reference standard was formed by averaging three specialists, reliability of the averaged coordinates was assessed separately for the x and y axes using a two-way random-effects, absolute-agreement, average-measures ICC. All verified specialist coordinates were retained.

Student tracing conditions

Each student completed two tracing conditions. Before beginning the tracing exercise, all students received a brief orientation to the tracing software and study procedure. No additional study-specific landmark-identification training or formal participant-calibration exercise was provided; students relied on their previous curricular teaching in cephalometric landmark identification.

In the unaided manual condition, students placed all eight anatomical landmarks without AI-generated landmark suggestions. The manual tracing files contained one sheet per student, with coordinate data for the included images and anatomical landmarks.

In the AI-assisted condition, students traced the same anatomical landmarks with access to AI-generated landmark suggestions. Students were permitted to accept the suggested position or adjust the landmark before saving the final tracing. The final saved landmark positions were used for analysis.

The analysis therefore compared students’ final landmark coordinates under unaided manual tracing and AI-assisted tracing against the same three-specialist reference standard.

Data extraction and cleaning

Annotation files were exported as spreadsheet or comma-separated files. Coordinate data were extracted from the annotation fields and converted into a long-format dataset containing image identifier, student identifier, tracing condition, landmark label, and x-y coordinate values.

All datasets were checked before analysis. Specialist files were checked to ensure that all specialists traced the same images and landmarks. Student files were checked to ensure that each student had complete data for both conditions. Landmark labels were harmonised across datasets, and only landmarks present in the reference standard and both student conditions were retained.

The final analysis dataset included 24 students, 11 cephalograms, and 8 anatomical landmarks per image, giving 2,112 landmark observations per student tracing condition.

Scale calibration

Errors were first calculated in pixels and then converted to millimetres using image-specific scale calibration. Scale calibration was derived from the scale landmarks recorded in the specialist reference files. The pixel distance between the two scale landmarks was measured for each image. The known 10 mm scale-bar length was then divided by this pixel distance to obtain the millimetre-per-pixel conversion factor.

This conversion factor was applied to all landmark errors for the corresponding image. The same calibration factor was used for both unaided manual tracing and AI-assisted tracing.

Landmark error calculation

For each student, image, landmark, and tracing condition, localisation error was calculated as the Euclidean distance between the student coordinate and the orthodontist-derived reference coordinate.

Error=(xstudent−xreference)2+(ystudent−yreference)2[1]

The resulting pixel error was converted to millimetres using the image-specific calibration factor.

Outcome measures

The primary outcome was MRE in millimetres. This was calculated overall, per student, per landmark, and per image for both unaided manual tracing and AI-assisted tracing.

Secondary outcomes included median error, SD, and SDRs within 2 and 4 mm. SDR within 2 mm was defined as the percentage of landmark observations with an error of 2 mm or less. SDR within 4 mm was defined as the percentage of landmark observations with an error of 4 mm or less.

Statistical analysis

The primary comparison was performed at student level because each student completed both tracing conditions. For each student, mean landmark error was calculated separately for unaided manual tracing and AI-assisted tracing. These paired student-level means were then compared.

The distribution of paired student-level differences was assessed using the Shapiro-Wilk test. Because the paired differences were non-normally distributed, the Wilcoxon signed-rank test was used as the primary inferential comparison. A paired t-test was also reported as a sensitivity analysis, together with a 95% confidence interval (CI) for the mean paired difference. The mean paired reduction was calculated as manual mean error minus AI-assisted mean error. Relative reduction was calculated as the mean paired reduction divided by the manual tracing mean error. Cohen’s dz was calculated as a standardised effect size for the paired comparison.

Per-landmark and per-image analyses were reported descriptively. Statistical significance was set at P<0.05.

Because the study used a counterbalanced crossover structure, students were grouped according to whether they completed unaided manual tracing first or AI-assisted tracing first. The mean reduction in landmark error was calculated separately for each sequence group, and the difference between groups was explored using Welch’s independent-samples t-test. This exploratory analysis assessed whether the observed association with AI assistance was consistent across tracing order.

An exploratory first-period sensitivity analysis compared student-level mean error among manual-first students during unaided manual tracing with that among AI-first students during AI-assisted tracing using the Mann-Whitney U test; Welch’s t-test and a 95% CI were additionally reported for the mean difference.

Missing data were not imputed; students were included only if complete coordinate data were available for both tracing conditions. Statistical analysis and data processing were performed using R software (version 4.4.3).

Educational framing

The study was designed to assess landmark localisation accuracy, but the comparison also has educational relevance. The unaided condition represented students’ independent landmark localisation after conventional teaching. The AI-assisted condition represented the same task when supported by AI-generated landmark suggestions. The difference between conditions was therefore used to assess the potential value of AI-assisted tracing as an educational scaffold in cephalometric training.

Code availability

The in-house AI-assisted tracing system was developed using Python and open-source libraries. The complete code package, including the training script, inference and metrics script, graphical application, configuration file, requirements file, and supporting source modules, is provided as Appendix 1. Patient images, identifiable data, and trained model weights are not included in the supplementary files. A local executable prototype was generated for demonstration and workflow purposes.

Ethical consideration

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Ethical approval for this study was obtained from the Institutional Ethics Committee of Oman Dental College (approval No. ODC\Res\2026\MSC005). The study used de-identified lateral cephalograms and student tracing data collected within an educational research context. Verbal informed consent was obtained from all student participants before they undertook the study procedures.

Results

Participant flow

Twenty-four eligible senior undergraduate dental students were enrolled and allocated to two counterbalanced tracing sequences. Twelve students completed unaided manual tracing first, followed by AI-assisted tracing, and 12 students completed AI-assisted tracing first, followed by unaided manual tracing. All 24 students completed both tracing conditions and were included in the final paired analysis. No student-level exclusions were applied after enrolment.

Dataset included in the analysis

The final analysis included 24 senior undergraduate dental students, 11 lateral cephalograms, and 8 anatomical landmarks per image. Each student completed both tracing conditions, producing 2,112 landmark observations for unaided manual tracing and 2,112 landmark observations for AI-assisted tracing. The reference standard was generated from the averaged coordinates of three orthodontic specialists for each image-landmark combination.

The analysed landmarks were S, N, A, B, UIT, UIA, LIT, and LIA. Scale landmarks were used only for image-specific calibration and were not included in the anatomical landmark accuracy analysis.

Inter-specialist reliability of the reference standard

Across 264 specialist–landmark observations (3 specialists × 11 images × 8 landmarks), the mean radial deviation from the corresponding three-specialist mean coordinate was 1.246 mm (SD 0.834). Landmark-specific mean deviations ranged from 1.128 mm for UIT to 1.403 mm for LIT, while maximum deviations ranged from 2.534 to 3.326 mm. The average-measures absolute-agreement ICCs for the calibrated x and y coordinates ranged from 0.974 to 0.995 across the eight landmarks, indicating excellent inter-specialist reliability (Table S1).

Technical implementation of the in-house AI-assisted tracing workflow

The in-house AI-assisted cephalometric tracing system was successfully implemented as a local deep-learning workflow and used to generate landmark suggestions for the AI-assisted tracing condition. During model development, 499 cephalograms were used for deep-learning training and internal evaluation, comprising 399 images for training, 50 images for validation, and 50 images for testing. The trained model was then applied to the 11 cephalograms included in the student-facing comparative evaluation.

The trained model produced predicted landmark positions for the included cephalograms, and the graphical application allowed users to view, adjust, and save final landmark coordinates. Complete AI-assisted coordinate data were obtained for all 24 students, 11 cephalograms, and 8 anatomical landmarks, indicating that the locally developed system could be incorporated into a student-facing tracing workflow for comparative evaluation.

Overall landmark localisation accuracy

AI-assisted tracing produced lower overall landmark localisation error than unaided manual tracing. Across all landmark observations, mean error was 3.169 mm (SD 2.869 mm) in the manual condition and 1.826 mm (SD 0.926 mm) in the AI-assisted condition. Median error was 2.320 and 1.733 mm, respectively. The SDR within 2 mm increased by 20.4 percentage points, from 42.5% to 62.9%, while the SDR within 4 mm increased by 33.1 percentage points, from 64.0% to 97.1% (Table 1).

Student-level paired comparison

At student level, mean landmark error was 3.169 mm with unaided manual tracing and 1.826 mm with AI-assisted tracing. The mean paired reduction was 1.343 mm (95% CI: 0.739–1.947), corresponding to a 42.4% relative reduction in mean error.

The paired differences were non-normally distributed (Shapiro-Wilk W=0.492, P<0.001). The primary Wilcoxon signed-rank comparison showed lower error with AI assistance (W=24, P<0.001); 22 of the 24 students had a lower mean error in the AI-assisted condition. The paired t-test sensitivity analysis gave t(23) =4.600 and P<0.001, with Cohen’s dz =0.939.

Per-landmark performance

Landmark-specific results were mixed. AI-assisted tracing had lower mean error for the four skeletal landmarks: S, N, A, and B, with mean differences of 3.994, 3.349, 2.034, and 3.022 mm, respectively. In contrast, unaided manual tracing had lower mean error for the four dental landmarks: UIT, UIA, LIT, and LIA, by 0.339, 0.208, 0.974, and 0.133 mm, respectively (Table 2). The magnitude and direction of these landmark-specific differences are illustrated in Figure 3.

Figure 3 Mean landmark localisation error by tracing condition. Open circles represent unaided manual tracing and filled blue circles represent AI-assisted tracing. Each point represents the mean of 24 student-level means, with each student mean calculated across 11 cephalograms. Connecting lines show the magnitude and direction of the condition difference for each landmark. Differences (Δ) were calculated as manual minus AI-assisted error; positive values indicate lower mean error with AI assistance, whereas negative values indicate lower mean error with manual tracing. AI, artificial intelligence.

In the AI-assisted condition, five landmarks had mean errors below 2 mm. A point, UIA, and LIA had the highest AI-assisted mean errors at 2.336, 2.139, and 2.053 mm, respectively.

The unmodified AI output comprised 88 unique predictions across the 11 cephalograms and eight anatomical landmarks. Raw AI predictions had a MRE of 1.431 mm (SD 1.157 mm) and a median error of 1.072 mm. The corresponding SDRs within 2 and 4 mm were 79.5% and 93.2%, respectively. Descriptively, the raw AI output had a lower mean and median error and a higher SDR within 2 mm than students’ final AI-assisted tracings, whereas the final student tracings had a higher SDR within 4 mm (97.1% versus 93.2%). At landmark level, students’ final AI-assisted positions had lower mean errors than the raw AI predictions for N, A, and UIA, while the raw AI predictions had lower mean errors for S, B, UIT, LIT, and LIA. These findings were treated as an exploratory descriptive comparison because raw AI output was not a separate participant tracing condition. Detailed landmark-specific descriptive comparisons between the raw AI output and students’ final AI-assisted positions are presented in Table S2.

Per-image performance

AI-assisted tracing produced lower mean error than unaided manual tracing in 10 of the 11 images. Manual mean error ranged from 1.850 to 3.989 mm, whereas AI-assisted mean error ranged from 1.594 to 2.191 mm.

The lowest AI-assisted mean error was observed for Ceph_7.jpg at 1.594 mm and the highest for Ceph_5.jpg at 2.191 mm. The lowest manual mean was observed for Ceph_2.jpg at 1.850 mm and the highest for Ceph_9.jpg at 3.989 mm. Ceph_2.jpg was the only image for which manual mean error was lower than AI-assisted mean error (1.850 versus 1.945 mm).

Effect of tracing sequence

The improvement associated with AI-assisted tracing was observed in both tracing sequences. Among students who completed unaided manual tracing first, mean error decreased from 3.200 mm in the manual condition to 1.816 mm in the AI-assisted condition, giving a mean reduction of 1.384 mm. Among students who completed AI-assisted tracing first, mean error decreased from 3.139 mm in the manual condition to 1.836 mm in the AI-assisted condition, giving a mean reduction of 1.302 mm.

No statistically significant difference in mean error reduction was found between the manual-first and AI-first sequence groups. The between-group difference in reduction was 0.081 mm (Welch’s test, P=0.89).

An exploratory first-period sensitivity analysis was restricted to observations obtained before crossover. Mean student-level error was 3.200 mm (SD 2.042 mm) among manual-first students during unaided manual tracing and 1.836 mm (SD 0.076 mm) among AI-first students during AI-assisted tracing. Error was lower in the AI-assisted group (Mann-Whitney U=120, exact P=0.005). The corresponding mean between-group difference was 1.364 mm [95% CI: 0.066–2.662; Welch’s t(11.03) =2.312, P=0.041], consistent in direction with the primary paired analysis.

Discussion

The present study found that AI-assisted cephalometric tracing substantially improved landmark localisation accuracy among senior undergraduate dental students compared with unaided manual tracing. Across all landmark observations, MRE decreased from 3.169 mm in the manual condition to 1.826 mm in the AI-assisted condition, while the SDR within 4 mm increased from 64.0% to 97.1%. AI-assisted mean error was lower for 22 of the 24 students and for 10 of the 11 images. However, the landmark-specific findings were mixed: AI-assisted mean error was lower for the four skeletal landmarks (S, N, A, and B), whereas unaided manual mean error was lower for the four dental landmarks (UIT, UIA, LIT, and LIA). No statistically significant difference in mean reduction was found between the tracing sequences (P=0.89). These findings support the role of AI-assisted cephalometric tracing as an educational scaffold that can improve overall student performance while preserving learner responsibility for verification and adjustment. They also demonstrate the importance of landmark-specific review. The excellent inter-specialist ICCs and the observed dispersion around the specialist mean support the reliability of the averaged reference standard used for these comparisons.

The magnitude of improvement observed in the AI-assisted condition is broadly consistent with the wider literature on automated cephalometric landmark detection. Recent systematic and review-level evidence indicates that AI-based landmark detection commonly achieves MREs within the range of approximately 1–2 mm, with clinically acceptable performance for many cephalometric landmarks (14-16). In the present study, the AI-assisted mean error of 1.826 mm and median error of 1.733 mm fall within this expected range, despite the fact that the analysed coordinates represented students’ final saved positions after interaction with AI-generated suggestions, rather than unmodified model outputs. This distinction is important because the study did not evaluate AI as a fully autonomous tracing substitute. Instead, it assessed how AI suggestions influenced learner-generated tracing accuracy within a supervised educational workflow.

The high SDR within 4 mm in the AI-assisted condition also aligns with previous reports showing strong performance of AI-assisted cephalometric systems at broader clinical thresholds (14,15). The higher SDR within 4 mm than within 2 mm is expected because the wider threshold includes observations with larger localisation deviations. Therefore, the high SDR within 4 mm should not be interpreted as equivalent to high precision within the tighter 2 mm margin. The SDR within 2 mm remained lower than the 4 mm threshold, indicating that AI assistance improved performance substantially but did not eliminate landmark-level variability. This pattern is consistent with previous work showing that some landmarks remain challenging even when AI is used, particularly points affected by anatomical overlap, reduced contrast, or ambiguity in radiographic definition (17,18). In the present study, A point, UIA, and LIA had the highest AI-assisted mean errors. Among these three landmarks, only A point had a lower AI-assisted mean error than manual mean error; unaided manual means were slightly lower for UIA and LIA.

The landmark-specific findings reinforce the view that AI-assisted cephalometry should be understood as a clinician-support and learner-support tool rather than as an autonomous replacement for human judgement. Previous studies evaluating automated and semi-automated cephalometric systems have shown that performance can vary by landmark, software platform, population, and image characteristics (17-19). Human verification remains particularly important for ambiguous landmarks and treatment-critical measurements. The overall reduction in error suggests that AI assistance may provide a useful starting point for students, while the mixed landmark-specific results demonstrate that every suggested point still requires active anatomical assessment and verification.

A descriptive comparison showed that the raw AI output had a lower overall mean and median error and a higher SDR within 2 mm than students’ final AI-assisted tracings, whereas the final student tracings achieved a higher SDR within 4 mm. Student review therefore appeared to reduce some of the larger localisation errors but did not improve the precision of AI-generated positions uniformly. Mean error was reduced after student interaction for N, A, and UIA but increased for the remaining five landmarks. These findings do not alter the primary comparison between unaided and AI-assisted student performance; rather, they demonstrate that a human-in-the-loop workflow should not be assumed automatically to improve every AI prediction. Educational implementation should therefore teach students to distinguish between justified correction of an inaccurate suggestion and unnecessary modification of an already accurate landmark position.

This interpretation is further supported by evidence on human-AI collaboration in cephalometric landmark detection. Studies examining collaborative workflows have reported benefits when AI-generated outputs are checked and, where appropriate, refined by human users, particularly when landmarks are difficult or images are less straightforward (20,21). The present study extends this concept into undergraduate dental education by showing that senior students, when supported by AI-generated landmark suggestions, achieved lower overall localisation error than when tracing unaided. This suggests that AI may provide a structured visual cue that guides attention without removing the need for judgement.

Clinical experience provides important context for interpreting landmark localisation. In a directly comparable educational study involving 20 dental students, unaided beginner tracing produced an overall MRE of 3.72 mm (SD 4.52 mm), with landmark-specific mean errors ranging from 1.31 to 8.90 mm, whereas beginner-AI collaboration produced an overall MRE of 3.14 mm (SD 4.06 mm) (21). The unaided MRE of 3.169 mm in the present study is therefore consistent with previously reported learner-level performance, although differences in the evaluated landmarks, images, and reference standards should be considered when making direct numerical comparisons. In a separate comparison involving an orthodontist, orthodontic resident, general dentist, and dental student, unaided cephalometric precision and accuracy generally followed experience level, whereas AI assistance produced larger improvements among the less-experienced examiners (22). Digital cephalometric training has also been shown to improve undergraduate landmarking accuracy (11). At the specialist level, automated performance has been reported to be broadly comparable with that of board-certified orthodontists for most evaluated landmarks, although discrepancies remained for selected points (17). The present findings fit this broader pattern: AI assistance substantially lowered overall error among senior students, although the mixed landmark-specific results indicate that it did not improve every landmark uniformly. AI may therefore provide less-experienced users with a structured starting point, while anatomical training and informed verification remain essential.

The educational implications are particularly important. Cephalometric tracing is not simply a technical drawing exercise; it requires students to interpret two-dimensional radiographs, recognise anatomical relationships, and apply landmark definitions accurately. Previous educational studies have reported that AI-assisted or digital cephalometric training can improve student landmarking accuracy, reduce tracing time, and support learner confidence (11,23,24). The present findings are consistent with this direction and suggest that AI-assisted tracing can be incorporated into preclinical or clinical orthodontic teaching as a formative learning tool. Within the broader landscape of dentomaxillofacial radiology education, such integration should be accompanied by deliberate curriculum development addressing AI literacy, critical appraisal of algorithmic outputs, and AI limitations, with educators taking an active role in guiding responsible use (25). A staged approach combining AI-supported practice with opportunities for unaided tracing and educator-led review may help students benefit from automated guidance while preserving independent image interpretation and anatomical reasoning, thereby limiting over-reliance.

A further contribution of this study is the use of an in-house AI-assisted tracing system rather than a commercial cephalometric platform. Much of the existing literature has evaluated externally developed or commercial systems, including automated landmarking platforms that may offer limited transparency regarding model design, training data, and deployment (17,18). In contrast, the present workflow was developed, trained, and deployed locally using open-source tools, then incorporated into a student-facing graphical application. This demonstrates the feasibility of local deep-learning development within a dental education context. Such an approach may have value beyond the immediate tracing task because it links clinical education with AI literacy, giving postgraduate trainees and educators exposure to data preparation, model training, validation, and practical implementation.

The in-house nature of the system should be interpreted carefully. Local development does not remove the need for validation, nor does it imply that the model is ready for unsupervised clinical use. Previous literature has emphasised that AI performance depends strongly on dataset quality, annotation consistency, population characteristics, and imaging device variability (19,26). The present study therefore supports feasibility and educational utility, but further work is needed before broader clinical deployment. Future studies should evaluate the model across larger and more diverse image sets, different malocclusion groups, and multiple imaging sources. External validation would be particularly important to determine whether the observed performance is maintained outside the local development environment.

The observation that AI-assisted tracing was associated with lower mean error in both sequence groups strengthens confidence that the direction of the finding was not confined to a single tracing order. Both the manual-first and AI-first groups showed reductions in error when AI assistance was available, and we did not find a statistically significant difference in mean reduction between the two sequence groups. This is important because crossover designs can be vulnerable to order effects, especially when the same task is repeated. Although the non-significant sequence comparison does not demonstrate equivalence, the similar direction of change in both groups supports the consistency of the overall association across tracing order under the tested conditions.

The interpretation of these findings should take into account several limitations. No a priori sample size calculation was performed. The student-facing evaluation used 11 cephalograms and 24 senior undergraduate dental students from a single educational context, which may limit external generalisability and statistical precision. The images were not stratified by malocclusion pattern or formally graded according to anatomical complexity, and participant-level data on previous orthodontic experience or academic performance were not collected. The same cephalograms were used in both tracing conditions; consequently, practice, carryover, and image-familiarity effects cannot be separated fully from the effect of AI assistance, even though tracing order was counterbalanced. The sequence comparison was also based on only 12 students per group and should be regarded as exploratory. Although the deep-learning model was developed using a larger image set, the educational evaluation itself was based on a smaller set of images. The study did not include repeat tracing under the same condition; therefore, intra-examiner test-retest reliability and limits of agreement could not be assessed. The study also focused on landmark localisation accuracy rather than downstream cephalometric measurements, treatment-planning decisions, or long-term retention of tracing skills.

Future research should examine whether repeated exposure to AI-assisted tracing improves unaided landmarking ability over time. Longitudinal studies could assess whether students who train with AI assistance retain better landmark identification skills when the AI scaffold is removed. Further work should also compare different levels of AI support, such as immediate feedback, delayed feedback, or blinded AI suggestions, to determine which approach best promotes learning while avoiding over-reliance. In line with broader recommendations for AI integration in dental and orthodontic education, future implementation should also include faculty development, clear governance, and explicit teaching on AI limitations, bias, and responsible use (7,27,28).

Conclusions

AI-assisted cephalometric tracing improved overall landmark localisation accuracy among senior undergraduate dental students compared with unaided manual tracing, although the magnitude and direction of the difference varied by landmark. The findings support further evaluation of a human-AI collaboration model in which AI provides landmark suggestions while students remain responsible for inspection, correction, and final placement. By demonstrating that an in-house deep-learning system can be incorporated into a student-facing tracing workflow, this study highlights the potential of locally developed AI tools as practical educational scaffolds in orthodontic training.

Acknowledgments

QuillBot was used to check grammar and punctuation.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0104/rc

Data Sharing Statement: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0104/dss

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0104/prf

Funding: None.

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0104/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Ethical approval for this study was obtained from the Institutional Ethics Committee of Oman Dental College (approval No. ODC\Res\2026\MSC005). The study used de-identified lateral cephalograms and student tracing data collected within an educational research context. Verbal informed consent was obtained from all student participants before they undertook the study procedures.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Subramanian AK, Chen Y, Almalki A, et al. Cephalometric Analysis in Orthodontics Using Artificial Intelligence-A Comprehensive Review. Biomed Res Int 2022;2022:1880113. [Crossref] [PubMed]
2. Dipalma G, Inchingolo AD, Inchingolo AM, et al. Artificial Intelligence and Its Clinical Applications in Orthodontics: A Systematic Review. Diagnostics (Basel) 2023;13:3677. [Crossref] [PubMed]
3. Kunz F, Stellzig-Eisenhauer A, Boldt J. Applications of artificial intelligence in orthodontics - an overview and perspective based on the current state of the art. Appl Sci 2023;13:3850.
4. Liu J, Zhang C, Shan Z. Application of Artificial Intelligence in Orthodontics: Current State and Future Perspectives. Healthcare (Basel) 2023;11:2760. [Crossref] [PubMed]
5. Arab A, Qutieshat A. Artificial intelligence in orthodontics: a narrative review of clinical and educational applications. J Med Artif Intell 2026;9:54.
6. Foroozandeh M, Salemi F, Shokri A, et al. Comparative accuracy of artificial intelligence-based AudaxCeph software, Dolphin software, and the manual technique for orthodontic landmark identification and tracing of lateral cephalograms. Imaging Sci Dent 2025;55:11-21. [Crossref] [PubMed]
7. Kunz F, Stellzig-Eisenhauer A, Widmaier LM, et al. Assessment of the quality of different commercial providers using artificial intelligence for automated cephalometric analysis compared to human orthodontic experts. J Orofac Orthop 2025;86:145-60. [Crossref] [PubMed]
8. Hanenkrath J, Park JH, Bay C. Training, use, and modifications related to artificial intelligence in postgraduate orthodontic programs in North America. Am J Orthod Dentofacial Orthop 2025;167:89-94.e2. [Crossref] [PubMed]
9. American Dental Association Standards Committee on Dental Informatics. ADA SCDI White Paper No. 1106 for Dentistry-Overview of Artificial and Augmented Intelligence Uses in Dentistry. Chicago: American Dental Association; 2022.
10. Nordblom NF, Büttner M, Schwendicke F. Artificial Intelligence in Orthodontics: Critical Review. J Dent Res 2024;103:577-84. [Crossref] [PubMed]
11. Lin J, Liao Z, Dai J, et al. Digital and artificial intelligence-assisted cephalometric training effectively enhanced students' landmarking accuracy in preclinical orthodontic education. BMC Oral Health 2025;25:623. [Crossref] [PubMed]
12. Hellyer P. Artificial intelligence in orthodontics. Br Dent J 2024;237:48. [Crossref] [PubMed]
13. Sharp C. Orthodontics in the city: artificial intelligence in orthodontics. BDJ Team 2025;12:168-71.
14. Ribas-Sabartés J, Sánchez-Molins M, d'Oliveira NG. The Accuracy of Algorithms Used by Artificial Intelligence in Cephalometric Points Detection: A Systematic Review. Bioengineering (Basel) 2024;11:1286. [Crossref] [PubMed]
15. de Queiroz Tavares Borges Mesquita G, Vieira WA, Vidigal MTC, et al. Artificial Intelligence for Detecting Cephalometric Landmarks: A Systematic Review and Meta-analysis. J Digit Imaging 2023;36:1158-79. [Crossref] [PubMed]
16. Rashmi S, Srinath S, Prashanth S, et al. Application of Artificial Intelligence in Automatic Cephalometric Landmark Annotations-A Comprehensive Review. In: International Conference on Advanced Communications and Machine Intelligence. Singapore: Springer Nature Singapore; 2023:167-82.
17. Ristau B, Coreil M, Chapple A, et al. Comparison of AudaxCeph®'s fully automated cephalometric tracing technology to a semi-automated approach by human examiners. Int Orthod 2022;20:100691. [Crossref] [PubMed]
18. Moreno M, Gebeile-Chauty S. Comparative study of two software for the detection of cephalometric landmarks by artificial intelligence. Orthod Fr 2022;93:41-61. [Article in French]. [Crossref] [PubMed]
19. Silva TP, Puntigam GS, Andrade-Bortoletto MFS, et al. Populational influence on cephalometric landmark identification: performance of two AI-driven software programs in Brazilian and Korean images. BMC Oral Health 2025;25:1596. [Crossref] [PubMed]
20. Davidovitch M, Sella-Tunis T, Abramovicz L, et al. Verification of convolutional neural network cephalometric landmark identification. Appl Sci 2022;12:12784.
21. Le VNT, Kang J, Oh IS, et al. Effectiveness of Human-Artificial Intelligence Collaboration in Cephalometric Landmark Detection. J Pers Med 2022;12:387. [Crossref] [PubMed]
22. Panesar S, Zhao A, Hollensbe E, et al. Precision and accuracy assessment of cephalometric analyses performed by deep learning artificial intelligence with and without human augmentation. Appl Sci 2023;13:6921.
23. Xu XX, Li X. The application of artificial intelligence-assisted tracing method in orthognathic cephalometric measurement teaching. China Journal of Oral and Maxillofacial Surgery 2025;23:72-6.
24. Alhazmi N, Alsaeed S, Almutairi L, et al. Accuracy of and dental students' preferences toward manual and digital cephalometric landmark identification: A randomized cross-over study. J Dent Educ 2024;88:815-22. [Crossref] [PubMed]
25. Negrete D, Lopes SLPC, Barretto MDA, et al. Artificial Intelligence and Dentomaxillofacial Radiology Education: Innovations and Perspectives. Dent J (Basel) 2025;13:245. [Crossref] [PubMed]
26. Lu G, Wang X, Xie M, et al. Bi-level alignment with super-resolution head for unsupervised cephalometric landmark localization. Phys Med Biol 2026. doi: 10.1088/1361-6560/ae35c5.
27. Ardila CM, Pineda-Vélez E, Vivares Builes AM. Integrating Artificial Intelligence into Orthodontic Education: A Systematic Review and Meta-Analysis of Clinical Teaching Application. J Clin Med 2025;14:5487. [Crossref] [PubMed]
28. Schwendicke F, Chaurasia A, Wiegand T, et al. Artificial intelligence for oral and dental healthcare: Core education curriculum. J Dent 2023;128:104363. [Crossref] [PubMed]