Applications of Artificial Intelligence in Dental Malocclusion: A Scoping Review of Recent Advances (2020–2025)

Abstract

Introduction: Dental malocclusion affects more than half of the global population, causing significant functional and esthetic consequences. The integration of artificial intelligence (AI) into orthodontic care for malocclusion has the potential to enhance diagnostic accuracy, treatment planning, and clinical efficiency. However, existing research remains fragmented, and recent advances have not been comprehensively synthesized. This scoping review aimed to map the current landscape of AI applications in dental malocclusion from 2020 to 2025. Methods: The review followed the Joanna Briggs Institute methodology and the PRISMA-ScR guidelines. The authors conducted a systematic search across four databases (PubMed, Scopus, Web of Science, and IEEE Xplore) to identify original, peer-reviewed research applying AI to malocclusion diagnosis, classification, treatment planning, or monitoring. The review screened, selected, and extracted data using predefined criteria. Results: Ninety-five studies met the inclusion criteria. The majority employed convolutional neural networks and deep learning models, particularly for diagnosis and classification tasks. Accuracy rates frequently exceeded 90%, with robust performance in cephalometric landmark detection, skeletal classification, and 3D segmentation. Most studies focused on Angle’s classification, while anterior open bite, crossbite/asymmetry, and soft tissue modeling were comparatively underrepresented. Although model performance was generally high, study limitations included small sample sizes, lack of external validation, and limited demographic diversity. Conclusions: AI offers the potential to support and enhance the diagnosis and management of malocclusion. However, to ensure safe and effective clinical adoption, future research must include reproducible reporting, rigorous external validation across sites/devices, and evaluation in diverse populations and real-world clinical workflows.

1. Introduction

Dental malocclusion, defined as the misalignment between the teeth of the two dental arches as the jaws close, is a clinically significant condition affecting individuals of all ages [1]. With an estimated global prevalence of over 50%, it contributes to esthetic concerns, functional impairments, and psychosocial distress, creating a growing demand for orthodontic care [2]. Malocclusion is associated with several health complications, including temporomandibular joint (TMJ) disorders, speech difficulties, and an elevated risk for dental caries and periodontal disease [3]. Traditionally, the diagnosis and management of malocclusion relied on clinical examination, cephalometric analysis, dental casts, photographic records, and orthodontic interpretation [4].

In recent years, artificial intelligence (AI) has demonstrated disruptive and transformative potential, particularly in complex pattern recognition and decision-making [5]. Dental and orthodontic providers are increasingly using AI to support and automate diagnostic workflows, enhance treatment planning, improve outcome prediction, and enable more precise monitoring [6]. By applying techniques such as machine learning, deep learning, and computer vision to radiographs, 3D scans, and photographic images, AI models can detect and classify malocclusion with potential gains in accuracy, efficiency, and reproducibility [7].

The application of AI in dentistry encompasses several technological domains. Machine learning enables computers to learn from data and make predictions or classifications without explicit programming [8]. Deep learning, a subset of machine learning, utilizes advanced neural networks such as Convolutional Neural Networks (CNNs), to analyze and interpret complex imaging data [9]. Computer vision seeks to replicate human visual perception in machines, enabling the automated analysis of visual dental inputs, including photographs and cephalometric images [10].

Despite the surge of interest and an increasing volume of published research, the field remains fragmented and poorly synthesized. Studies vary significantly in focus, AI techniques used, and malocclusion subtypes addressed [11]. There is a need for a comprehensive, updated synthesis of evidence to clarify AI’s status, identify which tools are truly effective, and outline the challenges that limit clinical translation.

A scoping review is well-suited for systematically mapping a rapidly evolving field across diverse study designs and clinical applications. Current studies span a wide range, from using machine learning to predict craniofacial growth [12], validating AI models for anterior open bite detection [13], and evaluating deep learning-based segmentation tools for orthodontic imaging [14]. However, clinicians often face difficulty when evaluating or adopting these tools due to inconsistent reporting and a lack of comparative validation.

Since 2020, there has been a marked acceleration in AI-driven orthodontic innovation, particularly in the use of deep learning and computer vision for tasks such as cephalometric landmark detection [15] and panoramic image classification for impacted canines [16]. Yet, the field lacks an overview of model reliability, clinical validity, and readiness for real-world integration.

This scoping review aims to fill this critical gap. Its objective is to explore the recent literature (2020–2025) on the use of AI in the diagnosis, classification, treatment planning, and monitoring of dental malocclusion. Specifically, it seeks to map and evaluate the technologies employed, the types and sources of input data, the clinical contexts of AI use, and report performance outcomes. In doing so, this review outlines prevailing trends, highlights areas of intensive development, and identifies research gaps that the field must address to move toward safe and effective clinical translation.

The remainder of this manuscript is organized as follows: Section 2 describes the scoping review methodology, including eligibility criteria, information sources, search strategy, screening, and data extraction. Section 3 presents the study selection results and synthesizes the included evidence across AI methods, clinical tasks, imaging modalities, and reported performance. Section 4 discusses clinical relevance, methodological trends, limitations, and research gaps to support future translation. Additionally, it provides conclusions and recommended directions for clinically reliable and generalizable orthodontic AI.

2. Methods

This scoping review followed the Joanna Briggs Institute methodology for scoping reviews, a structured approach for mapping existing evidence on a given topic [17]. The review also followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) checklist to ensure transparency and methodological clarity [18]. It was registered at the Open Science Framework (osf.io/6jqnd).

2.1. Eligibility Criteria

Table 1. Inclusion/exclusion criteria.

Inclusion Criteria	Exclusion Criteria
Human subjects’ dental applications Peer-reviewed original articles written in English Articles published between January 2020 and April 2025 Focused on AI and Dental Malocclusion	Conference proceedings Opinion articles (letter to the editor, editorial comments, opinion) Literature reviews (narrative, scoping, systematic, meta-analysis) Animal studies Abstracts without full text Case Reports

outlines the review criteria. The eligibility criteria included peer-reviewed original research published in English that explored the application of AI in the context of dental malocclusion, including diagnosis, classification, treatment planning, or monitoring. Studies were eligible if they implemented AI technologies (e.g., machine learning, deep learning, neural networks, or computer vision) to analyze or address malocclusion or related orthodontic applications. To be included, studies had to explicitly model malocclusion or an occlusal phenotype/trait (e.g., Angle class, vertical/transverse discrepancies, crowding/spacing, overjet/overbite) as an input, label, or endpoint, or evaluate an orthodontic decision relevant to malocclusion management (e.g., treatment planning/monitoring outcomes). Studies on adjacent conditions (e.g., TMD, airway) were included when malocclusion/occlusion traits were explicitly represented as model inputs and/or outputs, or when the endpoint was explicitly framed within orthodontic malocclusion assessment or treatment-planning pathways. The search excluded review articles, editorials, opinion pieces, case reports, abstracts, dissertations, and grey literature. To ensure the findings reflected current technological capabilities, the authors limited the search to studies published between January 2020 and April 2025.

We restricted inclusion to peer-reviewed original research to prioritize studies with sufficient methodological detail and evaluative reporting for meaningful synthesis. We limited studies to the English language to ensure consistent screening and interpretation of technical methods and outcomes. We selected the publication window (January 2020–April 2025) to reflect current AI capabilities and reporting practices, particularly the rapid expansion of deep learning and computer vision applications in orthodontics. We excluded reviews and other non-original publication types to avoid duplication of evidence and to focus the synthesis on primary model development and evaluation studies. Grey literature was excluded to improve comparability and because reporting of architectures, datasets, and validation procedures is often insufficient for extracting consistent methodological details.

For this review, we defined AI as computational methods that (i) learn patterns from data (machine learning/deep learning) or (ii) provide algorithmic decision support through explicit knowledge/rule representations (expert systems). Studies that were purely deterministic measurement tools without inference or decision logic were excluded. When commercial platforms did not disclose model architecture, we included them only if the study described an AI inference capability (e.g., automated detection/classification/segmentation or prediction) and reported a validation comparison.

2.2. Information Sources and Search Strategy

The review searched the literature across four major databases: PubMed, Scopus, Web of Science, and IEEE Xplore (

Table 2. Search strategies.

Database & Search Date	Search Strategy & Filters	Number of Studies Found
PubMed 14 April 2025	(“Artificial Intelligence” [Mesh] OR “Artificial Intelligence” [tiab] OR “AI” [tiab] OR “Machine Intelligence” [tiab] OR “Machine Learning” [tiab] OR “Computer Vision” [tiab] OR “Deep learning” [tiab]) AND (“Malocclusion” [Mesh] OR “Malocclusion” [tiab] OR “Occlusion” [tiab]) AND (“Dent*”) Filters: English & Publication Date	124
Scopus 14 April 2025	TITLE-ABS ((“Artificial Intelligence” OR “AI” OR “Machine Intelligence” OR “Machine Learning” OR “Computer Vision” OR “Deep learning”) AND (“Malocclusion” OR “Occlusion”)) AND ALL ((“Dent*”)) AND PUBYEAR > 2019 AND PUBYEAR < 2026 AND (LIMIT-TO (LANGUAGE, “English”)) Filters: English & Publication Date	197
Web of Science 14 April 2025	(TS = ((“Artificial Intelligence” OR “AI” OR “Machine Intelligence” OR “Machine Learning” OR “Computer Vision” OR “Deep learning”) AND (“Malocclusion” OR “Occlusion”))) AND ALL = ((“Dent*”)) Filters: English & Publication Date	112
IEEE Xplore 14 April 2025	(“Artificial Intelligence” OR “AI” OR “Machine Intelligence” OR “Machine Learning” OR “Computer Vision” OR “Deep learning”) AND (“Malocclusion” OR “Occlusion”) AND (“Dent*”) Filters: English & Publication Date	25

). The authors selected these databases to capture both the medical/dental literature and technical/engineering literature. Four authors (MH, OC, NB, CZ) collaborated with a research librarian (CS) using a combination of controlled vocabulary (e.g., MeSH terms in PubMed) and free-text keywords. The search terms targeted three main concepts: (1) AI (e.g., “artificial intelligence”, “machine learning”, “deep learning”, “neural networks”, “computer vision”), (2) malocclusion and related terms (e.g., “malocclusion”, “dental malocclusion”), and (3) imaging or data modalities (e.g., “radiograph”, “cephalometric”, “photograph”, “intraoral scan”). Boolean operators (“AND”, “OR”) were used to combine these terms appropriately. Table 2 outlines the search strategy.

To adapt the strategy across databases, we translated controlled vocabulary (e.g., MeSH in PubMed) into database-specific subject headings where applicable and adjusted field tags (e.g., TITLE-ABS, TS) and wildcarding. We iteratively tested synonym blocks for AI terms, malocclusion/occlusion terms, and dental/orthodontic context terms to ensure consistent recall across platforms (Table 2). Backward and forward citation tracking of included studies was also performed to reduce the risk of missed eligible studies.

2.3. Study Selection

All search results were imported into the EndNote reference management software Version 21, and duplicates were removed. Three reviewers (OC, NB, CZ) independently screened the titles and abstracts using the predetermined inclusion criteria. The same reviewers retrieved the full-text versions of the screened studies and independently assessed the articles for inclusion. The reviewers resolved inclusion disagreements through discussion, and if they could not achieve a consensus, a fourth reviewer (MH) made the final decision. Figure 1 summarizes the study selection process, including reasons for exclusion at the full-text screening stage.

2.4. Data Extraction

Three reviewers (OC, NB, CZ) independently extracted data from the included studies using a structured form developed in Microsoft Excel. The form included the following data items: authors, year of publication, country of origin, study design, data sources (e.g., radiographs, photographs, 3D scans), type of malocclusion addressed, specific AI methodology used, purpose of AI application (e.g., diagnosis, treatment planning), training and validation approach, and outcome measures (e.g., accuracy, sensitivity, specificity, F1 score, and study population). Additionally, the reviewers noted any reported limitations, ethical considerations, or practical challenges related to the AI application. To ensure accuracy and consistency, a fourth reviewer (MH) resolved discrepancies and conducted random quality checks on the extracted data.

2.5. Data Synthesis

The authors analyzed the extracted data using both quantitative and qualitative techniques. A descriptive numerical summary characterized the study distribution by year, country, AI type, and clinical application. Tabular presentations highlighted the types of AI models and malocclusion conditions studied. For the qualitative component, the authors conducted a thematic narrative synthesis, systematically reviewing the qualitative findings, identifying recurring concepts, grouping these concepts into broader themes, and then interpreting these themes in the context of the research aims. This approach identified emerging patterns, common AI applications, performance metrics, methodological trends, and key research gaps and underexplored areas. Given heterogeneity in reported metrics, datasets, and validation designs, performance findings were synthesized descriptively rather than pooled quantitatively.

3. Results

3.1. Overview of Study Selection

The search initially yielded 458 records across four databases: PubMed (124), Scopus (197), Web of Science (112), and IEEE Xplore (25). After removing 178 duplicate entries, 280 unique records remained for title and abstract screening. Abstract review excluded 174 additional studies. Reasons for exclusion included: 106 studies were unrelated to malocclusion, 30 were review articles or meta-analyses, 26 were not peer-reviewed original research, 10 were conference abstracts, and 2 were non-English, leaving 106 full-text reports. The full-text article review excluded an additional 10 articles; five were off-topic, two were conference abstracts, two were reviews, and one was a case report. Ultimately, the review identified 95 eligible research studies (Appendix A). The PRISMA-ScR flow diagram in Figure 1 outlines the full study selection process.

3.2. Publication Trends and Geographic Distribution

The number of publications increased over the review period, reflecting growing interest in the application of AI to dental malocclusion (

Table 3. Distribution of studies by year and country.

Years	Total Number of Studies	Top Contributing Countries: Number of Studies
2020–2021	16	South Korea: 5, Saudi Arabia: 1, Australia: 1, Taiwan: 1, Turkey: 1, USA: 1, Germany: 2, India: 1, Egypt: 1, UAE: 1, Slovakia: 1
2022–2023	26	China: 8, Saudi Arabia: 1, Germany: 2, South Korea: 3, Turkey: 1, Malaysia: 1, USA: 5, Iraq: 2, Denmark: 1, Australia: 1, Iran: 1
2024–2025	53	Egypt: 1, Saudi Arabia: 3, Iraq: 1, Brazil: 2, Iran: 1, Turkey: 2, China: 20, India: 2, Taiwan: 3, Australia: 2, Colombia: 1, Japan: 2, Sweden: 1, South Korea: 1, Israel: 1, Denmark: 1, USA: 2, UAE: 1, Thailand: 2, Italy: 2, Germany: 1, UK: 1

). The search identified 16 studies published between 2020 and 2021 and 26 studies from 2022 to 2023. The most significant surge occurred between 2024 and early 2025, with 53 studies published. This upward trend highlights the expanding research activity in this field. Geographically, China contributed the most with 28 studies, followed by South Korea (9), the USA (8), Germany (5), Saudi Arabia (5), Australia (4), Turkey (4), India (3), and Iraq (3) (Table 3).

3.3. AI Techniques Used in Malocclusion Research

The studies encompassed a wide range of AI methodologies, with convolutional neural networks (CNNs) being the most frequently used technique, featured in 36 studies (37.9%) (

Table 4. Artificial intelligence types and frequency.

AI Type	Number of Studies	% of Total
CNN	36	37.9%
Support Vector Machine, Random Forest	35	36.8%
Deep learning	15	15.8%
Hybrid/Ensemble methods	1	1.1%
Others	8	8.4%
Total	95	100.0%

). Thirty-five studies (36.8%) used traditional machine learning approaches, including support vector machines (SVM), random forest (RF), and k-Nearest neighbors. Other studies utilized deep learning models that did not fall strictly under CNN architectures (15 studies, 15.8%). Only 1.1% of studies employed hybrid or ensemble models, while 8 studies (8.4%) used alternative or unspecified AI techniques (Table 4).

3.4. Clinical Applications of AI

The diagnosis and classification of malocclusion was the most common clinical AI application, appearing in 55 studies (

Table 5. Clinical application of artificial intelligence in malocclusion studies.

Clinical Application	Number of Studies *	Common AI Models Used
Diagnosis/classification	55	Deep learning-based segmentation models Convolutional Neural Networks (CNNs) WebCeph software
Treatment planning	27	WebCeph software 3D Slicer software Deep learning with transformer architectures
Severity assessment	26	EfficientNetV2 and DenseNet201 Custom CNNs YOLO-based models
Growth prediction	3	Vision Transformer Self-supervised pre-training architectures Hybrid deep learning models

* Some studies contributed to more than one clinical application category (i.e., categories are not mutually exclusive), and some studies used more than one model.

). These studies focused on determining skeletal classes, detecting cephalometric landmarks, and identifying features such as overjet and overbite. Twenty-seven studies used AI for treatment planning, where it assisted in generating orthodontic or surgical recommendations based on diagnostic inputs. Twenty-six studies assessed severity, typically quantifying the degree of malocclusion or dental crowding. A smaller group of three studies applied AI to predict longitudinal craniofacial growth patterns, helping inform early intervention strategies. Several studies reported on specific tools or model families used for these applications, including WebCeph, 3D Slicer, and EfficientNetV2, reflecting the integration of proprietary and open-source platforms in AI-based orthodontic research. Because some studies addressed more than one clinical task, counts across application categories are not mutually exclusive.

3.5. Types of Malocclusion Conditions Studied

Angle’s classification system (Class I, II, and III) was the most frequently used malocclusion framework, referenced in 56 studies (

Table 6. Malocclusion condition studied.

Condition Category	Number of Studies *	Notes
Angle classification (I/II/III)	56	Includes any mention of Class I, II, or III malocclusion.
Anterior open bite	2	Specifically mentions of open bite.
Crossbite, crowding, asymmetry	10	Includes references to spacing, crowding, asymmetry, or crossbite.
Mixed/Other	30	Studies with general or unspecified malocclusion descriptions or where classification was unclear.

* Some studies addressed more than one condition.

). Other malocclusion types received more limited focus: just two studies [19,20] addressed anterior open bite, while 10 studies [21,22,23,24,25,26,27,28,29,30] explored crossbite, crowding, and asymmetry. Thirty studies either did not specify a malocclusion category or referred to mixed/generalized cases, indicating variability in diagnostic focus and inconsistent reporting across the literature. Because some studies addressed more than one condition, counts are not mutually exclusive.

3.6. Key Themes and Patterns Identified

A thematic synthesis revealed that the studies employed AI technologies to detect malocclusion types, classify treatment needs, analyze cephalometric landmarks, predict soft and hard tissue changes, segment dental structures, and support surgical planning (

Table 7. Summary of themes, performance, and gaps.

Category	Key Findings
Emerging patterns	Artificial intelligence (AI) models were used to detect malocclusion types, classify treatment needs, analyze cephalometric landmarks, predict soft and hard tissue changes, segment dental structures, assess facial symmetry, and assist in surgical planning [35]. Overjet, incisor changes, dental crowding, and skeletal discrepancies (especially Class III) appeared as targets of prediction or analysis [32]. AI supported early detection through profile images or photographs, and several studies focused specifically on Class III and deepbite malocclusions [36]. Some models integrated multiple modalities [14] (e.g., CBCT + profile photos), while others used only facial images or intraoral scans.
Performance metrics	Accuracy Most AI tools achieved 80–95% accuracy [37,38], with multiple models exceeding 90% for malocclusion classification, cephalometric tracing, and 3D segmentation [39,40]. Top accuracy results included: ResNet18: 96.5% (crossbite detection) [27] Inception V3: 92.6% [16] MobileNetV3 (with enhancement): 90.7% [26] 3D segmentation (Dice score): 95.1% [41] Malocclusion classification: ~93.1% (molars) [42] WebCeph agreement with clinicians: ICC = 0.90 Models analyzing hard tissue landmarks or simpler classification problems [15] were more accurate than those working with soft tissue or mixed cases [43]. Sensitivity & Specificity Sensitivity and specificity were generally high, especially for clear-cut skeletal Class III and crossbite cases. Examples include: ResNet18: Sensitivity 89%, Specificity 99.2% [27] Inception V3: Sensitivity 93.6%, Specificity 91.3% [16] TMD model: Sensitivity 71.2%, Specificity 84.4% [44] Class III screening from profile photos: Sensitivity 60%, Specificity 92.5% [45] Sensitivity was lower for soft tissue tasks or borderline cases [37], while specificity remained strong overall [46]. Precision Several models showed high precision, especially those with preprocessing or task-specific architectures [38]. Reported values include: ResNet18: 97.8% [38] Inception V3: 93.6% [16] MobileNetV3 (with CLAHE): top performer in precision among its group [26] Malocclusion (molars and canines): ~88.7% [42] TMD model (external test): F1 = 0.75 [44] Precision tended to be lower in ambiguous malocclusion types (e.g., Class I) [44] and higher for clearly defined skeletal cases [44]. Stability Many models demonstrated consistent performance across validation folds and test sets [47]. Studies using 5-fold or 10-fold cross-validation showed minimal performance variation [41]. WebCeph and DSS tools had stable agreement rates across subgroups [19]. Growth and symmetry models maintained accuracy across bootstrap or external validations [48,49]. Models using image enhancement (e.g., CLAHE) and hybrid training strategies tended to report better stability [26].
Overall Outcomes	Across studies, AI tools consistently supported accurate diagnosis, faster workflows, and reduced manual effort [12,50,51,52,53]. Many models outperformed or matched junior clinicians [24,54] Tools proved especially valuable in screening, cephalometric analysis, surgical planning, and underserved settings [33] Tasks involving 3D segmentation, skeletal Class III detection, and landmark tracing were the most successful [47,54,55,56,57,58,59,60,61,62,63] Performance declined in soft tissue, borderline cases, or when there was limited training data [64].
Methodological trends	Deep learning models such as ResNet18, MobileNetV3, Inception V3, and DenseNet were commonly used [26]. Many studies enhanced performance through image preprocessing (e.g., CLAHE), augmentation, and hybrid model design [25]. Interpretability was addressed by using GradCAM, SHAP, and VisualBackProp [47,65,66]. Data sources included cephalograms, CBCT, intraoral images, and profile photos [30,41,67,68,69]. Decision support tools (e.g., DSS, nomograms) were developed to translate AI predictions into actionable clinical guidance [19].
Evidence Gaps	Performance declined in borderline malocclusions (e.g., Class I), soft tissue predictions, and extraction decision modeling [60]. Studies often had small samples (e.g., 34–66 patients), limited to single centers, and lacked external validation [70,71,72,73,74]. Most models developed for specific ethnic groups (e.g., Chinese, Korean, Saudi), limiting their generalizability [48,75]. Manual landmarking required in many applications [76,77]. Soft tissue accuracy lagged hard tissue modeling [50]. Long-term outcome predictions rarely addressed.
Recommendations	Improvements should focus on better handling of borderline and soft tissue cases. Integration into clinical workflows is a priority. Improved interpretability tools will help gain clinician trust and enhance usability in everyday practice.

). Predictive targets frequently included overjet, incisor position, dental crowding, and Class III skeletal discrepancies. Three studies [31,32,33] used AI for early screening from profile photographs or intraoral images, particularly for Class III or deepbite conditions. Notably, a small subset of models [14,31,34] used multiple input modalities, such as CBCT and facial photographs, to enhance diagnostic performance. Overall, most studies focused on image-based diagnostic workflows, whereas fewer studies evaluated downstream outcomes (e.g., treatment planning impact or longitudinal prediction).

3.7. Performance of AI Models

Most AI models demonstrated robust performance across diagnostic and analytical tasks (Table 7). Reported accuracy typically ranged from 80% to 95%, with several models exceeding 90% in malocclusion classification, cephalometric landmark detection, and 3D segmentation [16,25,27,37,78,79]. For example, ResNet18 [27] achieved 96.5% accuracy in detecting crossbite, while Inception V3 [16] reached 92.6%. WebCeph [79] demonstrated strong agreement with clinicians, with an intraclass correlation coefficient (ICC) of 0.90.

Sensitivity and specificity scored high, especially for clearly defined conditions such as skeletal Class III and crossbite. One model [27] using ResNet18 reported 89% sensitivity and 99.2% specificity. In contrast, tasks involving soft tissue analysis or borderline cases had lower sensitivity, though specificity remained consistently high. Precision was also high for most models, with values such as 97.8% for ResNet18 and 93.6% for Inception V3 [16]. Although the studies did not consistently report F1 scores, the models that incorporated task-specific architectures and advanced preprocessing methods tended to report the highest-performing results.

To improve interpretability, we summarized performance descriptively and stratified reporting by imaging modality and by malocclusion category/clinical task, where sufficient information was provided. By imaging modality, performance reporting was most frequently derived from cephalograms and CBCT, with additional evidence from facial photographs and intraoral images; however, differences in acquisition protocols, devices, preprocessing pipelines, and reported metrics limited direct comparability across modalities. By malocclusion category/clinical task, performance evidence was strongest for common classification problems (e.g., Angle Class I–III and skeletal pattern classification), cephalometric landmark detection, and segmentation, whereas fewer studies reported performance for less common phenotypes and more complex tasks (e.g., soft tissue analysis and borderline cases). Accordingly, reported performance should be interpreted in the context of dataset composition, validation design, and the specific modality/task evaluated.

Cross-validation methods confirmed model stability. Studies using five- or ten-fold cross-validation showed low variance across folds. WebCeph demonstrated stable agreement rates across different demographic groups, and image enhancement techniques like CLAHE were linked to improved consistency and robustness. Nevertheless, heterogeneity in validation approaches and incomplete reporting of subgroup performance constrained quantitative synthesis and limited the ability to compare robustness across modalities and malocclusion subtypes.

3.8. Methodological Trends

Across the included studies, deep learning frameworks such as ResNet18, DenseNet201, Inception V3, and MobileNetV3 were the most widely used architectures (Table 7). Performance improvements were commonly attributed not only to architecture choice but also to optimization decisions, including preprocessing, augmentation, and hyperparameter selection. Preprocessing techniques (e.g., CLAHE), data augmentation, and hybrid learning strategies were frequently used to improve model performance. However, reporting of these optimization steps was heterogeneous, and many studies did not clearly describe whether tuning was prespecified, how hyperparameters were selected, or how experimental safeguards against data leakage were implemented (e.g., ensuring patient-level separation between training/validation/test sets and applying preprocessing within the training fold only). In robust experimental design, optimization procedures should be reproducible and systematically evaluated using a defined protocol (e.g., fixed data splits or nested cross-validation) rather than ad hoc trial-and-error. Structured optimization frameworks drawn from design-of-experiments, including parametric orthogonal-array approaches such as the Taguchi method [80], can reduce the search burden while enabling systematic, transparent tuning of preprocessing parameters and model hyperparameters.

The AI tools GradCAM, SHAP, and VisualBackProp addressed interpretability. Data sources included cephalograms, CBCT scans, intraoral images, and profile photographs, with several studies developing decision-support systems to provide clinically relevant outputs. Overall, clearer documentation of optimization workflows (preprocessing pipelines, hyperparameter search strategy, stopping criteria, and leakage controls) would improve reproducibility and strengthen confidence in reported performance gains.

3.9. Evidence Gaps and Limitations

Despite promising results, the review identified several gaps in evidence and limitations. AI models performed less accurately in borderline clinical cases, particularly Class I malocclusion, where diagnostic ambiguity reduced precision (Table 7). Performance also declined in soft tissue-related tasks, which remain challenging due to anatomical variability and lower contrast in imaging. Few studies addressed long-term treatment outcome prediction, limiting insights into the durability of AI-assisted planning. Methodologically, many studies relied on small and demographically homogeneous datasets, often from a single center, reducing generalizability. A widespread lack of external validation further restricted confidence in model robustness. Additionally, several models needed manual annotations of cephalometric landmarks, limiting scalability and increasing labor intensity.

4. Discussion

4.1. Summary of Findings

This review highlights a rapid acceleration in orthodontic AI research since 2020, reflecting growing enthusiasm for integrating automated tools into clinical practice [81]. CNNs and deep learning approaches dominated the methodological landscape, owing to their strong image-processing capabilities and adaptability. Most studies examined diagnostic and classification tasks involving Angle’s malocclusion classification (Class I, II, III), where AI models demonstrated high clinical relevance and technical accuracy. Reported accuracy rates frequently surpassed 90%, with tools such as CephNet, Cephio, and WebCeph achieving particularly strong performance in specialized tasks such as cephalometric landmark detection and 3D segmentation. These findings underscore the clinical potential of AI in improving diagnostic consistency, reducing manual effort, and expanding access to orthodontic care. However, because performance metrics and validation strategies were inconsistently reported across studies, these high accuracies should be interpreted in light of study design, dataset size, and the presence (or absence) of external validation.

4.2. Comparison with Prior Reviews or Traditional Methods

Unlike previous reviews that broadly surveyed AI use in dentistry or included mixed study types, this scoping review synthesized research specifically designed for malocclusion applications [82,83,84]. By narrowing the search to studies published between 2020 and 2025, this review captured recent advancements in deep learning, machine learning, and computer vision applied to orthodontics. The exclusion of reviews, abstracts, and grey literature enhanced methodological consistency and clinical relevance. This narrower scope enabled a more precise mapping of AI trends, tools, and knowledge gaps, offering new insights that extend earlier, more generalized analyses.

Compared to traditional manual approaches [85], such as cephalometric tracing and skeletal classification by orthodontists, AI methods offer substantial advantages in speed, scalability, and reproducibility [86,87,88,89,90]. Deep learning models, particularly CNNs, can process radiographs and 3D scans rapidly and provide consistent outputs without fatigue or inter-operator variability [19,56,91,92]. Accuracy matched or exceeded junior clinicians, particularly in landmark detection and skeletal Class III classification [54]. Nonetheless, clinical translation depends not only on accuracy but also on reproducibility, generalizability across sites/devices, and integration into real-world workflows.

However, significant challenges remain. Model interpretability is often lacking, making it difficult for clinicians to trust or understand AI-generated results. Generalizability is another key concern, as models may perform well in internal validation but struggle when applied to external datasets, different imaging protocols, or different devices. Taken together, these limitations highlight the need for transparent reporting, rigorous external validation, and clinically grounded implementation pathways before widespread adoption.

4.3. Methodological Considerations in the Reviewed Studies

A methodological strength across the reviewed studies was the consistent use of advanced deep learning models, including architectures like ResNet, DenseNet, and MobileNet, which enabled high performance in image analysis tasks. Several studies also explored more clinically realistic workflows by integrating multiple imaging modalities (e.g., combining cephalograms, CBCT, and facial photographs), which may better reflect clinical decision-making and case complexity [78].

However, notable limitations were observed. Many studies relied on small, single-center datasets with limited demographic diversity, reducing the generalizability of their findings. Most studies failed to perform external validation, raising questions about a model’s robustness in real-world practice. Furthermore, workflows often required manual preprocessing steps, such as image cropping, landmark annotation, or quality filtering, limiting their full automation and scalability in clinical settings.

Methodological heterogeneity remains a key limitation. Studies varied widely in reporting of architectures, dataset composition, train/validation/test splitting strategies, and metrics, making direct comparisons difficult and limiting reproducibility. To address this, we propose a minimal reporting checklist tailored to orthodontic AI research (

Table 8. Minimum reporting checklist for orthodontic AI and related clinical AI studies.

Checklist Domain	Minimum Item to Report	Examples/Notes
Clinical aim	Intended use and clinical task	Screening vs. diagnosis vs. planning vs. monitoring; target malocclusion traits.
Population	Eligibility criteria and phenotype definitions	Define Angle class and other traits (overjet/overbite, crowding, asymmetry).
Data provenance	Source, centers, dates, and sample size	Single vs. multi-center; number of patients/images; missing data handling.
Demographics	Age/sex/ethnicity distribution	Report and justify; note representativeness of target population.
Imaging protocol	Device/protocol metadata and acquisition settings	Modality, vendor/model, geometry, resolution; exposure where applicable.
Reference standard	Labeling process and annotator expertise	Who labeled, guidelines used; inter-/intra-rater reliability.
Preprocessing	Preprocessing and augmentation pipeline	Normalization/CLAHE; augmentation types; applied only on training folds.
Data splitting	Train/validation/test strategy and leakage controls	Patient-level split; nested cross-validation when tuning; no overlap across splits.
Model details	Architecture and training configuration	Backbone, loss, optimizer, epochs, batch size; initialization/transfer learning.
Hyperparameters	Tuning strategy and search space	Grid/random/Bayesian; stopping rules; design of experiment/Taguchi as systematic option.
Evaluation	Primary metrics with uncertainty	Task-appropriate metrics, confidence intervals; calibration for probabilistic outputs.
External validity	External and cross-domain validation	Testing across sites/devices; domain shift analysis; harmonization/adaptation methods.
Subgroup analysis	Phenotype- and demographic-stratified performance	Rare phenotypes; fairness across demographics; error analysis and failure modes.
Interpretability & safety	Explanation method and clinical use	Grad-CAM/SHAP; how explanations are presented; safety monitoring plan.
Clinical workflow	Prospective workflow/user evaluation	Usability, chair time, acceptance; impact on decisions; reporting of adverse events.

), adapted from established AI reporting frameworks (CONSORT-AI [93], TRIPOD-AI [94]) and imaging-focused guidance (CLAIM [95]), emphasizing the minimum information needed to interpret validity, bias, and clinical readiness. Although Table 8 is presented as an orthodontic-AI checklist, its domains (e.g., reference standard, data provenance, external validation, subgroup performance, and workflow evaluation) reflect broadly applicable principles for trustworthy clinical AI. As such, the checklist can be readily adapted to other areas of dentistry and broader healthcare, as well as to related applied AI settings where transparency, generalizability, and safety-oriented evaluation are required.

Terminology was also used inconsistently across the literature. To support clearer future reporting, we recommend a standardized taxonomy distinguishing: (i) machine learning (non-neural models), (ii) deep learning (neural models), (iii) CNNs (deep learning subtype for images), (iv) transformers/ViTs (deep learning subtype), (v) hybrid models (e.g., two-stage, multimodal fusion, or physics-informed approaches), and (vi) decision support systems (learning-based vs. rule-based). Standardizing definitions would reduce ambiguity in synthesis and improve cross-study interpretability. We also recommend that future studies explicitly label the model family and pipeline type in the Methods (e.g., ML vs. DL; CNN vs. transformer; unimodal vs. multimodal; end-to-end vs. two-stage) to support clearer synthesis and reduce inconsistent terminology across publications.

4.4. Clinical Implications

AI is poised to reshape orthodontic workflows by enabling faster, more consistent, and more objective diagnostic processes. Models trained on cephalograms, CBCT scans, and clinical photographs can produce rapid assessments, reduce clinician workload, and minimize diagnostic variability. Tools like WebCeph are already in use, offering real-time support for automated tracing, classification, and treatment simulation.

Importantly, AI applications show promise for remote screening, particularly beneficial in underserved or rural areas with limited access to orthodontic specialists. However, clinical reliability may be reduced for underrepresented malocclusion phenotypes. Only a small number of studies addressed anterior open bite, crossbite, or facial asymmetry, which increases the risk of spectrum bias and limits confidence when these models are applied to less common presentations. Because most studies did not report phenotype-stratified performance (e.g., rare phenotype sensitivity/specificity, confidence intervals, and error modes), this review could not quantify the impact of sparse phenotype coverage on model bias; nevertheless, the clinical implication is clear: Models trained predominantly on common phenotypes may underperform for rare or complex cases, potentially increasing misclassification risk and inappropriate treatment planning.

Interpretability is also a practical clinical need rather than a purely technical preference. Explanations can support error detection in borderline cases, calibrate clinician trust, document rationale for accountability, and facilitate shared decision-making with patients. Interpretability methods do not require a “special needs interpreter”; rather, they require clinically designed interfaces and appropriate user training so explanations are understandable, standardized, and not misleading.

Finally, although several studies proposed decision support systems, few evaluated real-world outcomes such as workflow impact, chair time, usability, or user acceptance. Future work should include prospective workflow studies assessing operational benefit and clinical safety, including clinician interaction, failure handling, and user experience.

4.5. Limitations

This scoping review followed established methodological guidance, which does not require a formal quality appraisal of included studies. However, while this approach prioritizes breadth over depth and effectively captures the scope, diversity, and evolving nature of AI applications in malocclusion research, the absence of a structured risk-of-bias assessment warrants caution when interpreting the strength of individual study findings and the certainty of reported performance claims.

Limiting the review to English-language publications between January 2020 and April 2025 may miss relevant studies in other languages or published before 2020. Nonetheless, the review deliberately chose this period to focus on the most current and technologically advanced developments, particularly the rapid rise in deep learning and computer vision tools in clinical orthodontics.

Finally, while most studies offer insight into AI performance, some lacked comprehensive reporting on technical aspects such as model architecture, training parameters, or validation design. This limited the ability to perform direct methodological comparisons. Most importantly, limited reporting of subgroup performance (by phenotype and demographics) and limited external validation constrained our ability to quantify bias, fairness, and real-world generalizability across clinical settings and patient populations. Despite these limitations, the review provides a timely and meaningful synthesis that lays the groundwork for future focused evaluations, interdisciplinary research, and clinical translation.

4.6. Research Gaps and Future Directions

Future research should address several notable gaps identified in the current literature. Certain malocclusion types, such as anterior open bite, crossbite, and facial asymmetry, remain underrepresented. Addressing these gaps will require deliberate sampling strategies and routine reporting of phenotype-specific errors, particularly in borderline cases and less common subtypes.

AI performance also remains suboptimally studied in borderline cases, particularly Class I malocclusion, and in soft tissue modeling, where anatomical variability and imaging limitations introduce additional challenges. Soft-tissue-related tasks frequently appear less robust than hard-tissue tasks, and demographic homogeneity (ethnicity, age, sex) combined with geographic concentration of datasets likely worsens generalizability. Because demographic stratification and subgroup performance were often not reported, we could not quantify the effect; however, future studies should routinely report subgroup performance and calibration, especially for soft-tissue endpoints.

Very few studies explored long-term treatment outcomes, an area crucial for advancing personalized orthodontic care. Future “long-term prediction” studies should prioritize clinically meaningful outcomes such as relapse/stability (alignment and occlusal indices), retention outcomes, long-term skeletal and soft tissue changes, stability of planned tooth movement, and patient-centered outcomes where available.

A major gap across the literature is the lack of external validation. Practical strategies to improve generalizability include multi-center external testing, cross-device and cross-protocol evaluation, harmonization of imaging pipelines, domain adaptation approaches, federated learning across institutions, preregistration of evaluation protocols, and curated reporting of device/protocol metadata to enable domain-shift analysis. In addition, hybrid simulation–AI approaches [96] (e.g., combining physics-based modeling with AI) can support controlled generation of synthetic data to augment underrepresented conditions, reduce overfitting, and strengthen external validity testing (consistent with emerging examples in the broader biomedical AI literature). In imaging-based orthodontic AI, cross-device and cross-protocol testing is a clinically essential form of external validation because equipment and acquisition differences can produce meaningful domain shift.

Most studies still rely on unimodal inputs (single imaging modality). Future research should more explicitly pursue multimodal fusion approaches (e.g., combining cephalograms, CBCT, facial photographs, and intraoral scans) to better reflect clinical decision-making and improve model robustness and clinical usefulness.

Finally, future work should focus on developing interpretable AI systems that integrate seamlessly into clinical workflows, enabling clinicians to act on AI-generated recommendations with greater confidence. Clinical translation will require not only technical validation but also prospective evaluation in real-world settings, including usability, chair time reduction, clinician acceptance, and safety-oriented reporting consistent with early-stage clinical evaluation guidance for decision support tools. These studies are essential to demonstrate real-world value beyond retrospective performance metrics and to support responsible clinical adoption.

4.7. Conclusions

AI holds significant promise for enhancing the diagnosis, classification, and treatment planning of dental malocclusion. With high accuracy and the potential to streamline clinical workflows, AI tools can augment orthodontic care by increasing diagnostic consistency, reducing manual workload, and expanding access to underserved populations. However, to realize this potential, future efforts must prioritize robust external validation, model interpretability, and seamless integration into clinical settings. To support clinical translation, future work must move beyond high internal accuracy toward reproducible reporting, rigorous external validation across sites and devices, explicit evaluation of rare phenotypes and demographic subgroups, and prospective workflow trials demonstrating real-world clinical value. In addition, future studies should adopt standardized terminology and minimum reporting practices to enable meaningful comparison across datasets, devices, and clinical contexts. Ultimately, demonstrating clinical readiness will require prospective evaluation in representative populations, transparent documentation of performance limitations, and evidence that AI tools improve decision-making and efficiency without compromising patient safety. Advancing the field will require close collaboration between clinicians, data scientists, and engineers to ensure that AI systems are both technically sound and useful in real-world orthodontic practice.