Cephalometric analysis involves identifying anatomic landmarks from lateral skull radiographs and calculating linear and angular measurements to evaluate craniofacial morphology. Cephalometric analysis traditionally requires the identification of specific anatomic landmarks of the craniofacial complex. This identification requires a detailed knowledge of anatomy with precision for accuracy in diagnosis, and the subsequent analysis of the craniofacial morphology of the patient is dependent on a high level of accuracy and precision. Historically, the process is inherently susceptible to human errors, including intraobserver and interobserver clinician variability alongside technical errors, all of which can potentially lead to inaccuracies in cephalometric tracing and diagnosis.

The development of CNNs has shown high accuracy in the identification of cephalometric landmark, feature extraction, and patient diagnosis. The architecture of CNNs is specifically designed to process visual data with a grid-like structure, enabling CNNs to learn and prioritize patterns crucial for a given task. Unlike traditional ML methods that require manual selection of relevant features, CNNs can automatically learn which patterns are important by extracting raw features from image data. It is important to note that CNNs still require expert-labeled training data, such as annotated landmarks on cephalometric radiographs, to learn the desired predictions. However, once trained, they exhibit robust generalization capabilities without manual input. Cephalometric analysis has seen success with CNNs, an ML technique particularly effective at analyzing visual data, automated landmark detection has significantly improved the speed and consistency of cephalometric tracing ( Fig. 1 ).

CNN-based cephalometric landmark detection pipeline: model prediction compared to expert ground truth. CNN, convolutional neural network.

DL methods were found to achieve better results than examiners for the detection of certain cephalometric landmarks, especially those locates on anatomic curves. The CNN-based architecture of the YOLOv3 (You-Only-Look-Once) model for object detection has demonstrated especially strong performance, with its architecture designed to predict bounding boxes and class probabilities simultaneously in a single forward pass through the network; this makes the model highly efficient for real-time image analysis while maintaining special precision. Studies show that YOLOv3 and PANN (Pretrained Audio Neural Network) models found equivalent results from AI-based detection as compared to human experts. ^, Reported success detection rates (SDRs) ranging from 75% to 90% at a 2 mm threshold and some advanced methodologies even achieve accuracies at high as 98.2% SDR at 2 mm with test data mean radial error of 0.61 mm on facial profile images in controlled settings ( Table 2 ).

Three-dimensional imaging has become a cornerstone of modern orthodontic diagnosis, with the rise of cone-beam computed tomography (CBCT) and intraoral scanning providing comprehensive volumetric data of both soft and hard tissues. Although 3D data enable more comprehensive evaluations, the analysis of the images may be time consuming and prone to variability. AI, particularly with DL, has shown strong potential in automating 3D analysis and offering a more streamlined approach to diagnosis and treatment planning.

CBCT imaging provides volumetric representations of the craniofacial complex. To extract clinically useful information from these volumetric datasets, preprocessing a CBCT, specifically segmentation, is essential to allow orthodontists to isolate and visualize specific anatomic structures from raw CBCT images. Segmentation involves delineating individual structures like dental crowns, dental roots, airway, and other relevant anatomy. Traditionally, this process was performed manually or with semiautomated tools embedded into imaging software (eg, edge detection, watershed segmentation, clustering methods). However, achieving accurate anatomic segmentation remains challenging, requiring laborious manual correction in clinical settings to achieve accuracy, especially in the field of orthodontics where the presence of metal artifacts, like metal brackets, can distort the image and complicate interpretation.

Traditionally, this process was performed manually or with semiautomated tools embedded into imaging software (eg, edge detection, watershed segmentation, clustering methods). However, achieving accurate anatomic segmentation remains challenging, requiring laborious manual correction in clinical settings to achieve accuracy, especially in the field of orthodontics where the presence of metal artifacts, like metal brackets, can distort the image and complicate interpretation.

To address these limitations, researchers have developed CNNs capable of identifying and segmenting individual teeth with orthodontic brackets directly from raw CBCT images. One specific model was able to segment and identify all teeth in about 14 seconds with an Intersection over Union score of 0.99, indicating that the AI’s predicted tooth boundaries overlapped 99% with the expert-segmented tooth area, correctly classifying each tooth within the mouth. Notably, the model was slightly less accurate when segmenting anterior tooth compared with posterior teeth, likely explained from their smaller size and variable morphology.

Intraoral scanners have become a standard tool within orthodontics, generating highly accurate digital models of the dentition and surrounding soft tissue. Similar to CBCTs, they require segmentation and classification of individual teeth and relevant structures for diagnostic value. DL methods for segmenting intraoral scans have shown to outperform manual human segmentation in speed and efficiency; manual segmentation was reported to take from a few minutes to 2 hours as compared with DL methods that took 4.2 seconds to 5 minutes with a mean accuracy of 94.26% with a 95% CI of 92.67% to 95.85%. AI-based intraoral scan analysis supports downstream applications like automated arch form recognition, occlusion assessment, and tracking of treatment progress, especially in aligner treatments, that can greatly aid the clinician ( Fig. 2 ).

Traditional versus AI-based diagnostic workflows.

The increased accessibility of 3D visualization facilitated by AI has greatly enhanced the orthodontist’s ability to visualize complex anatomic relationships and more effectively treatment plan cases. Clinicians can now identify previously obscured pathology, such as lateral incisor root resorption due to ectopically erupting canines, with greater ease and accuracy. AI-enabled tools now make it possible to virtually removal of overlying teeth or structures from 3D films, providing a clearer view of the underlying anatomy. These capabilities open new dimensions in orthodontic treatment planning, a deeper understanding of individual patient anatomy leading to improved diagnostic precision and more informed clinical decision making.

Craniofacial classification is a fundamental component of orthodontic diagnostics, traditionally based on lateral cephalometric analysis using angular and linear measurements (eg, ANB angle, Wits appraisal) to identify skeletal patterns (class I, II, III) and vertical proportions (eg, hyperdivergent or hypodivergent growth). These assessments, however, are subject to variability and require significant expertise; diagnostic errors in classification can misguide treatment.

ML models (eg, neural networks, support vector machine [SVM], logistic regression, XGBoost) have enabled classification of skeletal patterns with reported accuracies of over 90%. Deep convolutional neural network (DCNN) models have achieved an overall accuracy of 96% in skeletal classification, with class-specific accuracies of 97% (class I), 96% (class II), and 95% (class III). Emerging applications include automated facial image analysis simply from extraoral photographs to identify skeletal patterns. These nonradiographic alternatives serve as valuable tools for preliminary screening and triaging, especially in radiation-sensitive environments. ^,

Assessing craniofacial growth is critical to accurate diagnosis and treatment timing, as skeletal maturity influences orthodontic decisions around orthopedic appliances or orthognathic intervention. Orthodontists have historically used a variety of maturity indicators, such as stages of sexual development, chronologic age, and radiographic evaluation of ossification centers, like as those found in hand-wrist radiographs. To minimize radiation exposure, many clinicians rely on lateral cephalograms to assess cervical vertebral maturation (CVM) as a proxy for skeletal age.

AI models, especially those using customized DCNNs, have been applied to automate CVM assessment based on morphology from the cervical vertebrae C2 to C4 in the lateral cephalograms. These models have shown promising results, achieving 76% accuracy on unseen data. However, notable misclassifications persist, particularly in intermediate staged with significant underestimations for CVMS 4 and CVMS 5, whereas early (CVMS 1) and late (CVMS 6) stages exhibited minimal misclassifications. Although the technology for skeletal developmental assessment with AI remains in the early stages, ongoing improvements in model architecture with more training data offer significant potential in enabling diagnostic consistency and reducing clinician burden.

AI enables clinicians to leverage large datasets of previous treatment modalities and outcomes to make data-driven predictions about craniofacial development; using AI-based recommendations, they can adapt their treatment accordingly. ML algorithms have been developed to predict tooth movement trajectories and suggest personalized treatment plans. This fundamentally shifts the orthodontic paradigm from reactive correction to personalized, proactive prevention to enable clinicians to optimize treatment timing and focus on patient-centered outcomes that promote long-term oral health.

In tandem, predictive models have been developed to estimate orthodontic treatment duration. A recent study evaluating Random Forest, Lasso, and Elastic Net demonstrated comparable performance, each achieving a mean absolute error of approximately 7.27 months in predicting total treatment duration. Although current variability still limits exact case-by-case case prediction, these models hold clinical utility for setting realistic patient expectations, improving scheduling efficiency, and identifying outliers in treatment duration. As input variables and datasets continue to expand in size and diversity and become more refined, model predictive accuracy is expected to further improve.

AI models have also shown early success predicting the need for dental extractions during orthodontic treatment-based variables including gender, diagnostic models, and cephalometric records. One study, using automated ML software (Auto-WEKA), reported a 93.9% accuracy in predicting the need for orthodontic extraction using both radiographic and model variable inputs. Other studies found a 94% accuracy in AI-driven models predicting the need for extractions versus nonextraction treatment and 84.2% accuracy in determining extraction patterns. These models identified several key predictive factors including amount of crowding, anterior tooth positioning, lower incisor proclination, overjet, overbite, and lip competence. However, these models did not consider dental findings such as periodontal health, previous endodontic treatment, missing teeth, and heavily restored teeth, which remain essential components of comprehensive clinical judgment. Therefore, these models are not designed to dictate treatment decisions, but rather to assist clinicians in providing data-driven insights that can support and enhance their clinical judgment.