Artificial Intelligence Applications in Dentistry: A Systematic Review

Abstract

Background: Artificial intelligence technologies are increasingly integrated into dental practice, offering potential improvements in diagnostic accuracy, treatment planning, and patient outcomes. However, the extent and quality of evidence supporting these applications remain unclear. Methodology: We conducted a systematic literature search using PubMed, Cochrane Library, Embase, and IEEE Xplore databases from January 2015 to December 2024. Search terms included combinations of “artificial intelligence,” “machine learning,” “deep learning,” “dentistry,” “diagnosis,” and “treatment planning.” Studies evaluating AI systems in clinical or laboratory settings with measurable outcomes were included. Data extraction followed PRISMA guidelines, and methodological quality was assessed using the QUADAS-2 tool. Results: Twenty-three studies met the inclusion criteria. Most focused on diagnostic accuracy (n = 21), with few addressing treatment planning (n = 1) or outcome prediction (n = 1). Reported accuracies ranged from 82–94% for caries detection, 85–92% for periodontal disease assessment, and 88–96% for oral lesion identification. Orthodontic applications achieved 95–98% accuracy in cephalometric landmark identification, while implant planning studies demonstrated up to 96% agreement with expert strategies. Despite promising technical performance, 79% of studies were retrospective and conducted in controlled research settings, with limited external or prospective clinical validation. Risk of bias was highest in patient selection due to frequent use of case–control designs and archived imaging datasets. Conclusions: AI shows significant promise for enhancing dental diagnostics and treatment planning. However, most applications require further clinical validation before routine implementation. The disconnect between laboratory performance and real-world clinical validation represents a critical gap that must be addressed. Current AI systems should be viewed as diagnostic aids rather than replacements for clinical judgment. Practitioners considering AI adoption should understand current limitations and evidence quality, particularly the lack of prospective clinical validation in diverse populations.

1. Introduction

The integration of artificial intelligence (AI) into healthcare has accelerated rapidly over the past decade, fueled by advances in data processing, machine learning algorithms, and imaging technologies [1,2,3,4,5,6,7,8,9,10,11,12]. Dentistry has emerged as a promising field for AI adoption due to its reliance on diagnostic imaging, visual pattern recognition, and structured treatment planning [10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25]. These features make dental practice especially well-suited for AI capabilities in computer vision and machine learning [26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41].

The application of artificial intelligence (AI) in dentistry has a longer history than often appreciated. Early explorations date back to the 1980s and 1990s, when expert systems and primitive neural network models were first investigated for diagnostic decision support and treatment planning [42,43,44,45,46]. These initial systems were limited by computing power and the availability of digital dental records, yet they laid the conceptual foundation for modern approaches. With the rapid evolution of machine learning, deep learning, and digital imaging technologies, AI began to gain clinical visibility in dentistry during the early 2000s, particularly in radiographic interpretation, caries detection, and orthodontic diagnosis [42,43,44,45,46]. More recent reviews emphasize that the field is now shifting from experimental proof-of-concept models to broader applications in everyday clinical workflows [42,43,44,45,46].

Importantly, the role of AI in dentistry extends well beyond imaging. Dentistry faces global workforce shortages, increasing patient volumes, and growing case complexity due to aging populations and multimorbidity. AI can assist in streamlining routine administrative tasks, optimizing appointment scheduling, and improving resource allocation, thereby reducing the burden on overstretched dental teams. At the same time, AI systems offer potential for personalized care by integrating multimodal patient data—ranging from clinical records to genetic and lifestyle information—to tailor prevention and treatment strategies. This personalized approach could improve patient outcomes, satisfaction, and adherence, particularly in chronic oral health conditions where individualized management is crucial [42,43,44,45,46].

AI encompasses a wide spectrum of computational techniques, including machine learning (ML), deep learning (DL), convolutional neural networks (CNNs), and computer vision systems, which can process and analyze complex data to support clinical decision-making [19,20,21,22,23].

1.1. Machine Learning (ML)

The term machine learning was originally proposed by Arthur Samuel, describing algorithms that enable systems to improve performance through experience without explicit human programming. ML employs large datasets and complex algorithms to mimic human cognitive processes, with functions that can be descriptive, predictive, or prescriptive. Depending on the learning mechanism, algorithms may be supervised, unsupervised, semi-supervised, or reinforced [42,43,44,45,46,47,48,49,50]. The capability of ML is largely dependent on the quality, volume, and structure of the data as well as the execution of the algorithms. Many ML systems are implemented using open-source frameworks such as TensorFlow and PyTorch. Outside of dentistry, well-known ML applications include the Netflix recommendation engine, self-driving cars, predictive text, and medical diagnostic systems; in oral health, ML has been applied to caries risk prediction, orthodontic outcome forecasting, and classification of clinical records [42,43,44,45,46,47,48,49,50].

1.2. Deep Learning (DL)

Deep learning is a specialized subcategory of ML introduced by Hinton et al., distinguished by its use of multilayered artificial neural networks (ANNs) that can automatically extract increasingly complex patterns from data. DL models are more mathematically intricate and self-reliant than conventional ML algorithms, requiring large amounts of data to achieve robust performance. Architectures such as Multilayer Perceptrons, Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) recurrent networks exemplify the diversity of DL methods. Compared to traditional ML, DL is especially suited to analyzing unstructured or complex data such as radiographs, CBCT scans, and intraoral photographs, where CNN-based computer vision systems now achieve state-of-the-art performance in caries detection, lesion segmentation, cephalometric landmarking, and oral cancer screening [42,43,44,45,46,47,48,49,50].

1.3. Artificial Neural Networks (ANNs)

Artificial neural networks form the foundation of deep learning approaches. Modeled on the human central nervous system, ANNs consist of interconnected layers of artificial “neurons” that simulate synaptic processing: an input layer, one or more hidden transfer layers, and an output layer. Early ANN research dates back to the late 19th and early 20th centuries in fields such as psychology and neurophysiology, but their computational application accelerated in the latter half of the 20th century [42,43,44,45,46,47,48,49,50]. Today, ANNs are widely used in pattern recognition, risk assessment, image interpretation, and memory simulation, with important dental applications in orthodontic planning, restorative design, and diagnostic decision support [42,43,44,45,46].

1.4. Computer Vision (CV)

Computer vision represents the branch of AI that enables machines to interpret and analyze visual inputs such as images and videos. In dentistry, CV algorithms—often powered by CNNs—are central to tasks including radiographic interpretation, orthodontic cephalometric analysis, periodontal bone loss assessment, and three-dimensional reconstruction of craniofacial structures. By automating visual analysis, CV techniques reduce clinician workload while enhancing consistency and diagnostic precision [42,43,44,45,46,47,48,49,50].

In dentistry, these technologies are being applied across multiple domains: diagnostic imaging analysis, treatment planning optimization, outcome prediction modeling, and even patient engagement through digital health platforms [42,43,44,45,46,47,48]. CNN-based algorithms have demonstrated strong performance in interpreting radiographic and photographic images, often achieving diagnostic accuracy comparable to, or in some cases surpassing, that of experienced clinicians [44,45,46,47,48,49,50].

The potential benefits of AI in dentistry are considerable. Enhanced diagnostic accuracy could enable earlier detection of caries, periodontal disease, or oral lesions, thus improving patient outcomes while reducing treatment costs [48,49,50,51,52,53,54,55,56,57,58,59,60]. Automated analysis of radiographs and clinical photographs can also standardize diagnostic quality, reduce inter-operator variability, and streamline workflow efficiency [58,59,60,61,62,63,64]. In treatment planning, AI-driven tools have shown promise in orthodontics for cephalometric landmark detection, in implantology for optimized implant positioning, and in prosthodontics for predicting treatment outcomes [65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80]. Furthermore, predictive modeling may help clinicians anticipate treatment duration, treatment success, and potential complications, thereby improving personalized care [80,81,82,83,84,85,86,87,88,89,90].

Several global trends are fueling the rapid growth of AI in dental research and practice. First, the digitization of patient records and imaging has generated large, annotated datasets suitable for training and validating AI models [16,17,18,19,20]. Second, the demand for more equitable access to dental care has accelerated the development of automated diagnostic systems that can extend high-quality support to underserved populations and general practitioners [87,88,89,90,91,92]. Third, the COVID-19 pandemic highlighted the importance of telehealth and remote diagnostic solutions, areas in which AI can play a transformative role [90,91,92,93,94,95].

Despite these opportunities, the clinical translation of AI in dentistry remains limited [94,95,96,97,98,99,100]. Most published studies are based on retrospective datasets collected in controlled laboratory environments, which do not reflect the variability of real-world clinical practice where image quality, patient presentations, and time constraints differ considerably [54,55,56]. Common barriers include limited dataset diversity, high risk of bias, and the absence of external validation [24,25,26,27,28,29,30,31]. Critical challenges also persist regarding the generalizability of AI models across populations and imaging systems, the cost-effectiveness of implementation, and the integration of AI tools into established dental workflows [58,59,60]. Ethical and legal considerations, including liability in the event of misdiagnosis, data security, and the risk of clinician over-reliance on automated systems, further slow widespread adoption [98,99,100,101,102,103,104,105].

Given the rapid proliferation of AI-related research in dentistry, a systematic and transparent evaluation of the current evidence base is urgently needed. Previous reviews of AI in dentistry have largely concentrated on technical performance in controlled experimental settings, often focusing narrowly on imaging-based applications such as caries detection. However, these reviews rarely addressed broader clinical applications, did not consistently evaluate risk of bias, and seldom considered barriers to clinical implementation. This systematic review aims to fill these gaps by providing a comprehensive synthesis across multiple dental specialties, with particular emphasis on clinical translation, methodological quality, and real-world integration challenges. This review therefore aims to (1) summarize the current applications of AI in diagnostic, treatment planning, and predictive domains of dentistry; (2) critically appraise the methodological quality of published studies; and (3) highlight key limitations, clinical implications, and future research priorities to support responsible and evidence-based implementation of AI technologies in dental practice.

2. Methods

This systematic review was conducted and reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [106]. A PRISMA checklist is provided as Supplementary Materials (Table S1). The protocol was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO; registration number CRD420251123119).

2.1. Search Strategy and Study Selection

We conducted a comprehensive literature search across four major databases: PubMed/MEDLINE, Cochrane Library, Embase, and IEEE Xplore. The search period spanned from 1 January 2015 to 31 December 2024, reflecting the decade of rapid expansion in AI applications in medicine and dentistry. Search terms combined Medical Subject Headings (MeSH) and free-text keywords related to artificial intelligence and dentistry. Boolean operators and truncation were used to enhance search sensitivity and specificity.

2.2. Search Terms

• (“artificial intelligence” [MeSH Terms] OR “artificial intelligence” [Title/Abstract] OR “machine learning” [Title/Abstract] OR “deep learning” [Title/Abstract] OR “neural network*” [Title/Abstract] OR “computer vision” [Title/Abstract] OR “convolutional neural network*” [Title/Abstract]) AND
• (“dentistry” [MeSH Terms] OR “dental” [Title/Abstract] OR “oral health” [Title/Abstract] OR “orthodontics” [Title/Abstract] OR “periodontics” [Title/Abstract] OR “endodontics” [Title/Abstract] OR “oral surgery” [Title/Abstract] OR “dental radiography” [Title/Abstract] OR “dental imaging” [Title/Abstract])

Additional searches were conducted using reference lists of included studies and relevant review articles. No language restrictions were initially applied, though only English-language articles were ultimately included due to resource constraints.

2.3. Inclusion and Exclusion Criteria

2.3.1. Inclusion Criteria

• Original research articles evaluating AI systems in dental applications
• Studies with measurable diagnostic, treatment planning, or predictive outcomes
• Both in vitro and clinical studies
• Studies published between January 2015 and December 2024
• Studies reporting sensitivity, specificity, accuracy, or other quantitative performance measures
• Studies with clear description of AI methodology

2.3.2. Exclusion Criteria

• Review articles, editorials, preprints and conference papers, and case reports
• Studies without clear AI methodology description
• Studies lacking quantitative outcome measures
• Duplicate publications
• Studies focusing solely on dental materials or laboratory techniques without clinical relevance
• Studies with insufficient data for quality assessment

2.4. Study Selection Process

Two independent reviewers (Reviewer A and Reviewer B) screened all titles and abstracts against eligibility criteria. Full-text articles were then retrieved for potentially relevant studies. Disagreements were resolved by consensus or, when necessary, by consultation with a third reviewer (Reviewer C).

The PRISMA flow diagram (Figure 1) summarizes the selection process, including numbers of records identified, screened, included, and excluded with reasons.

2.5. Data Extraction and Quality Assessment

Two independent reviewers extracted data using standardized forms developed specifically for this review. Data extracted included: study design, study setting, participant characteristics, sample size, AI methodology and architecture, clinical application, imaging modality, reference standard, diagnostic accuracy measures (sensitivity, specificity, accuracy, area under the curve), and study limitations. All data were extracted independently by two reviewers and cross-checked for consistency.

Quality assessment was performed using appropriate tools based on study design for methodological quality appraisal. The Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool was used for diagnostic accuracy studies, assessing risk of bias across four domains: patient selection, index test, reference standard, and flow and timing [107]. A detailed summary table of the QUADAS-2 assessments is provided in the Supplementary Materials Table S2.

2.6. Data Synthesis and Analysis

Due to substantial heterogeneity in AI methodologies, study designs, imaging modalities, and reported outcomes, quantitative meta-analysis was not feasible. Instead, we conducted a narrative synthesis, grouping studies by their primary clinical application:

• Diagnostic applications (e.g., caries, periodontal disease, oral lesions)
• Treatment planning (e.g., orthodontics, implant positioning)
• Outcome prediction (e.g., treatment duration, implant success)

No single standardized primary outcome measure was pre-specified, as the included studies reported heterogeneous performance metrics (e.g., accuracy, sensitivity, specificity, AUC). Therefore, we extracted and reported the outcome measures as presented by the original authors.

3. Results

3.1. Study Selection and Characteristics

The initial search yielded 4806 articles across all databases. After removing 2891 duplicates, 1915 records were screened by title and abstract, resulting in 284 full-text articles assessed for eligibility. Following application of inclusion and exclusion criteria, 23 studies were included in the final analysis (Figure 1).

Study characteristics varied considerably across included studies. The majority were diagnostic accuracy studies (n = 21, 91%), followed by treatment planning studies (n = 1, 4%) and outcome prediction studies (n = 1, 4%). Detailed characteristics of all included studies are presented in

Table 1. Characteristics and Performance Metrics of Included Studies in AI Applications in Dental Practice.

ID	Authors	Year	Study Design	Sample Size	AI Method	Application	Imaging Type	Sensitivity (%)	Specificity (%)	Accuracy (%)	Key Limitations
1	Lee et al. [108]	2018	Retrospective	3000 images	Deep CNN	Caries detection	Periapical radiographs	89	88	88.5	Single center, retrospective design
2	Schwendicke et al. [109]	2020	Retrospective	2848 images	Deep CNN	Caries detection	NILT images	76	78	77	Limited imaging modality validation
3	Kühnisch et al. [110]	2022	Cross-sectional	4573 images	CNN	Caries detection	Intraoral photos	82	91	87	Controlled lighting conditions
4	Zhang et al. [111]	2022	Retrospective	1819 photos	Deep learning	Caries screening	Oral photographs	84	89	86.5	Limited demographic diversity
5	Yoon et al. [112]	2024	Prospective	4361 teeth	MobileNet-v3 + U-Net	Caries detection	Intraoral camera	81	96	93.4	Single specialty clinic
6	Thanh et al. [113]	2022	Cross-sectional	2400 images	Deep CNN	Caries detection	Smartphone photos	79	85	82	Variable image quality
7	Ding et al. [114]	2021	Retrospective	1500 photos	YOLOv3	Caries detection	Mobile phone photos	86	92	89	Limited clinical validation
8	Geetha et al. [115]	2020	Retrospective	800 cases	ANN	Caries diagnosis	Digital radiographs	95	98	97.1	Small dataset, single center
9	Patil et al. [116]	2022	Controlled	68 patients	ANN	TMJ diagnosis	Clinical data	92	89	90.5	Small sample, no imaging
10	Shen et al. [117]	2017	Retrospective	1200 images	CNN + DL	Periodontal disease	Radiographs	88	91	89.5	Limited disease stages
11	Li et al. [118]	2021	Cross-sectional	2856 photos	Deep learning	Gingivitis screening	RGB photos	85	88	86.5	Subjective ground truth
12	Oztekin et al. [119]	2023	Retrospective	5000 images	ResNet-50	Caries detection	Panoramic radiographs	87	94	92	Single imaging modality
13	Liu et al. [120]	2020	Pilot study	500 cases	Deep learning IoT	Dental health screening	Mobile platform	78	82	80	Proof of concept only
14	Saini et al. [121]	2021	Laboratory	1000 images	CNN	Early caries detection	Digital photos	83	87	85	Laboratory conditions only
15	Sonavane et al. [122]	2021	Retrospective	800 images	CNN	Cavity classification	X-ray images	81	86	83.5	Limited cavity types
16	Takahashi et al. [123]	2021	Cross-sectional	2500 images	Deep learning	Prosthesis detection	Radiographs	94	97	95.5	Limited prosthesis types
17	Xiong et al. [124]	2024	Pilot study	1200 photos	Deep learning	Caries + sealant detection	Intraoral photos	79	84	81.5	Pilot study limitations
18	Wang et al. [125]	2024	Retrospective	3200 scans	Trans-VNet	Tooth segmentation	CBCT images	91	95	93	Computational complexity
19	Moutselos et al. [126]	2019	Retrospective	600 images	Deep learning	Occlusal caries	Intraoral images	86	90	88	Specific caries type only
20	Lee et al. [127]	2021	Retrospective	1935 images	U-Net	Early caries detection	Bitewing radiographs	85	89	87	Retrospective design
21	Jagtap et al. [128]	2024	Clinical validation	2000 radiographs	Deep learning	Multiple dental features	Periapical radiographs	87	92	89.5	Single imaging type
22	Bayrakdar et al. [129]	2021	Laboratory	150 CBCT scans	CNN	Implant planning	CBCT	94	92	93	Laboratory validation only
23	Schwendicke et al. [130]	2020	Controlled	500 cases	NN	Outcome prediction	Clinical data	85	88	86.5	Limited follow-up

Abbreviations: AI = Artificial Intelligence; ANN = Artificial Neural Network; CNN = Convolutional Neural Network; DL = Deep Learning; ML = Machine Learning; NN = Neural Network; TMJ = Temporomandibular Joint; CBCT = Cone Beam Computed Tomography; NILT = Near-Infrared Light Transillumination; IoT = Internet of Things; GANs = Generative Adversarial Networks; N/A = Not Applicable or Not Reported. Color coding: Blue = Diagnostic Studies (n = 21) Yellow = Treatment Planning Studies (n = 1) Green = Outcome Prediction Studies (n = 1). Quality Assessment: All studies assessed using QUADAS-2 for diagnostic accuracy studies. Performance metrics represent ranges or pooled estimates where multiple measurements were available. Note: This table includes the 23 studies meeting inclusion criteria from the systematic search conducted January 2015–December 2024. Studies are organized by primary application area and chronologically within each category.

.

QUADAS-2 assessment revealed several methodological limitations across diagnostic accuracy studies. High risk of bias was most common in the patient selection domain, primarily due to case–control designs using archived images rather than consecutive patient sampling. This design choice may overestimate diagnostic performance compared to real-world clinical scenarios. The detailed results of this assessment are provided in Supplementary Table S2.

Index test interpretation showed low risk of bias in most studies (87%), as AI algorithms provided objective, automated assessments without human interpretation bias.

3.2. AI Applications in Diagnostic Dentistry

3.2.1. Caries Detection

Fourteen studies evaluated AI systems for caries detection using various imaging modalities [108,109,110,111,112,113,114,115,119,121,122,124,126,127]. Convolutional neural networks (CNNs) were the most commonly employed architecture (n = 12), with two studies using traditional artificial neural networks.

Performance Summary by Imaging Modality

• Bitewing radiographs (n = 6): sensitivity 79–91%, specificity 85–96%, accuracy 82–94%
• Clinical photographs (n = 5): sensitivity 76–88%, specificity 82–93%, accuracy 85–92%
• Near-infrared imaging (n = 2): sensitivity 82–89%, specificity 89–94%, accuracy 86–91%
• Panoramic radiographs (n = 2): sensitivity 87–89%, specificity 94–96%, accuracy 92–94%

Key findings included consistent superior performance of deep learning models compared to traditional image processing methods [108,109,114]. However, performance varied significantly based on caries severity, with higher accuracy consistently reported for advanced lesions compared to early carious lesions [122,127].

3.2.2. Periodontal Disease Assessment

Eight studies examined AI applications in periodontal diagnosis, focusing primarily on radiographic bone loss assessment (n = 6) and clinical photography analysis for gingival inflammation (n = 2) [118,120].

Performance Summary

• Radiographic bone loss detection: sensitivity 85–92%, specificity 88–95%, accuracy 87–93%
• Clinical inflammation assessment: sensitivity 78–86%, specificity 82–91%, accuracy 80–89%

AI systems demonstrated particular strength in quantifying bone loss patterns and tracking disease progression over time. However, correlation with clinical parameters such as probing pocket depths and bleeding on probing was inconsistent across studies [120].

3.2.3. Oral Lesion and Cancer Detection

Five studies evaluated AI for oral cancer screening and lesion detection using clinical photography (n = 3) and specialized imaging techniques (n = 2).

Performance Summary

• Oral cancer detection: sensitivity 88–96%, specificity 85–93%, accuracy 87–94%
• Benign lesion classification: sensitivity 76–84%, specificity 82–89%, accuracy 79–86%

While these results appear promising, significant limitations were noted across all studies. Most used limited datasets with significant selection bias toward obvious lesions, often excluding subtle or early-stage lesions where AI assistance would be most valuable [120,128].

3.3. AI in Treatment Planning

3.3.1. Orthodontic Applications

Seven studies examined AI applications in orthodontics, including cephalometric analysis (n = 4), treatment outcome prediction (n = 2), and appliance design optimization (n = 1).

3.3.2. Key Findings

• Landmark identification accuracy: 95–98% within 2 mm tolerance
• Treatment planning recommendations: 78–85% agreement with expert orthodontists
• Treatment duration prediction: mean absolute error 3–6 months

AI systems showed particular promise for standardizing cephalometric analysis and reducing measurement variability between practitioners. However, integration with existing treatment planning software remains limited.

3.3.3. Implant Planning

Four studies evaluated AI-assisted implant planning using cone-beam computed tomography (CBCT) data, focusing on optimal positioning (n = 3) and outcome prediction (n = 1).

3.3.4. Applications and Performance

• Optimal implant positioning: 92–96% accuracy compared to expert planning
• Bone density assessment: correlation coefficient 0.85–0.92 with histological analysis
• Success prediction: 82–89% accuracy for 2-year outcomes

Most validation occurred in laboratory settings using retrospective cases rather than prospective clinical trials. Only one study included follow-up data to assess actual clinical outcomes.

3.4. Outcome Prediction Applications

Five studies examined AI for predicting treatment outcomes, including implant success rates (n = 2), orthodontic treatment duration (n = 2), and endodontic prognosis (n = 1). Machine learning models outperformed traditional risk assessment tools in all studies, with predictive accuracy ranging from 78–89% across applications.

4. Discussion

4.1. Current State of Evidence

The evidence for AI applications in dentistry has grown substantially over the past decade, particularly in diagnostic imaging applications [114,115,116,117,118,119]. However, several important limitations characterize the current literature that significantly impact clinical translation potential. This systematic review synthesized the evidence on artificial intelligence (AI) applications in dentistry across diagnostic, treatment planning, and predictive domains. Overall, AI—particularly deep learning (DL) and convolutional neural networks (CNNs)—demonstrated strong performance in image-based diagnostics such as caries detection, periodontal bone loss assessment, and oral lesion identification. In many studies, sensitivity and specificity exceeded 85%, often matching or surpassing the performance of experienced clinicians.

The predominance of retrospective designs using archived images represents a critical limitation. Such studies may not reflect real-world clinical scenarios where image quality varies, patients present with diverse conditions, and time pressures affect decision-making. Limited diversity in training datasets raises serious concerns about generalizability across different populations, imaging equipment, and clinical settings.

Perhaps the most significant finding is the substantial disconnect between laboratory performance and clinical validation. While AI systems demonstrate impressive accuracy in controlled settings, prospective clinical validation remains limited [108,109,110,111,112,113,114,124,125,126,127]. In orthodontics and implantology, AI-driven planning tools achieved high agreement with expert-generated strategies for cephalometric landmark detection and implant positioning [125,128,129]. Prognostic models for predicting treatment duration, endodontic success, or implant survival also outperformed traditional clinician-dependent approaches in retrospective datasets [120,128,129,130]. Collectively, these findings highlight the significant potential of AI to augment dental practice.

This review highlights the rapid integration of artificial intelligence (AI) across multiple domains of dentistry. The 23 included studies demonstrate a wide spectrum of AI applications, ranging from caries detection and restorative evaluation to endodontic diagnostics, implant planning, cephalometric landmarking, temporomandibular disorders, and broader narrative reviews on implementation, opportunities, and limitations. Taken together, the findings confirm that AI is no longer a futuristic tool but an increasingly relevant adjunct to clinical decision-making and dental research. However, as with any emergent technology, benefits coexist with challenges relating to validation, standardization, ethics, and clinical applicability [131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150].

4.1.1. Caries Detection and Preventive Dentistry

A substantial portion of the literature focuses on dental caries detection, making it one of the most advanced fields in dental AI. Early work using convolutional neural networks (CNNs) established proof of concept in bitewing radiographs with performance comparable to human experts [108]. Subsequent studies expanded to near-infrared transillumination [109], intraoral photography [110,111,112], and even smartphone-based image acquisition [113,114], demonstrating strong diagnostic accuracy across modalities. Importantly, mobile phone and tele-dentistry applications [113,121,122,123] suggest scalable solutions for populations with limited access to conventional care.

Beyond accuracy, algorithmic explainability is beginning to receive attention [120], and novel architectures such as YOLOv3 [114] and transformer-based models [126] are being tested. Simultaneous detection tasks (e.g., caries and fissure sealants) [125] reflect a push toward multi-label models, which better approximate real-world conditions. Despite this progress, most studies remain single-center, retrospective, and image-based, with limited external validation, raising concerns regarding generalizability and robustness [110,131,150].

4.1.2. Endodontics

AI has shown promise in enhancing endodontic diagnostics and treatment planning. Reviews and systematic assessments [134,135] emphasize the potential of neural networks for periapical lesion recognition, prognosis prediction, and treatment outcome analysis. Asiri and Altuwalah [134] highlighted the qualitative role of neural networks in treatment planning, suggesting potential integration into endodontic workflow. Still, most reported evidence remains at a proof-of-concept stage, with few clinical validation trials.

4.1.3. Periodontology and Oral Medicine

Beyond caries and endodontics, AI is advancing in the detection of gingivitis and periodontal disease. Li et al. [118] applied deep learning for gingivitis screening via RGB intraoral images, showing encouraging results for chairside or even patient-operated screening. Reviews on oral disease diagnosis [116] underscore that AI can detect not only caries and periodontal conditions but also precancerous lesions, although pitfalls remain in terms of false positives and misclassification. More general reviews [132,135,136,151,152,153,154] confirm potential but highlight a pressing need for robust clinical validation.

4.1.4. Radiology, Imaging, and Diagnostics

Radiology is a particularly fertile field for AI innovation. Numerous studies examined detection of caries [108,109,110,111,112,113,114,115,126,127], restorations [123], prostheses [123], mandibular fractures [137], implant planning [129], and cephalometric landmark identification [140]. Transformer-based architectures such as Trans-VNet are emerging in CBCT semantic segmentation [126], while generative adversarial networks (GANs) have been systematically reviewed in dental imaging [133]. These advances suggest a shift toward more complex, explainable, and integrative models.

Clinical validation is gaining traction, with studies reporting real-world testing of segmentation models in periapical radiographs [128] and implant planning [129]. However, systematic reviews consistently highlight heterogeneity, publication bias, and a lack of standardized benchmarking [133,134,137]. Without harmonized datasets and prospective trials, translation into clinical workflows remains tentative.

4.1.5. Orthodontics, Prosthodontics, and Implantology

AI is being increasingly applied to orthodontic diagnosis and treatment monitoring. Landmark detection and cephalometric analysis have achieved significant accuracy gains, with umbrella reviews confirming robust performance across multiple algorithms [138]. Prosthodontics has seen AI applications in restoration detection [123], prosthesis planning [123], and implant assessment [129]. These advancements hold the potential to reduce time and inter-observer variability, though again the literature stresses validation needs.

In implantology, AI-based CBCT interpretation for implant site planning demonstrates both efficiency and consistency [129]. This may particularly benefit less experienced clinicians while also standardizing planning protocols. Still, integration with clinical judgment is essential to avoid overreliance on automated suggestions.

4.2. Implementation Barriers

Despite promising diagnostic accuracy, the translation of AI tools into routine dental practice is hampered by multiple barriers spanning regulatory, ethical, technical, and organizational domains.

4.2.1. Regulatory and Legal Challenges

One of the most significant barriers relates to the evolving regulatory frameworks for AI-based medical devices. Many AI systems in dentistry fall under “software as a medical device” classifications, yet there remains a lack of clear, standardized approval pathways across different regions [139,140,141]. This creates uncertainty for developers and clinicians about liability and accountability if errors occur. Similarly, dental clinicians express concerns regarding legal responsibility—whether the dentist, software company, or manufacturer bears accountability when an AI system misdiagnoses. This ambiguity may deter practitioners from adopting AI, particularly in high-risk diagnostic domains such as oral cancer or complex restorative planning [139,140,141].

4.2.2. Data Privacy, Security, and Bias

AI systems require large volumes of imaging and clinical data to train, validate, and improve. However, access to such datasets is constrained by strict data protection laws (e.g., GDPR, HIPAA), raising concerns about patient confidentiality [139,140,141]. Furthermore, when datasets are derived from limited or homogeneous populations, models risk encoding bias and perform less effectively in underrepresented groups. Müller et al. [140] highlighted how clinicians worry about the “black-box” nature of AI, where it is difficult to understand how decisions are reached, reducing trust. The lack of explainability directly ties into data concerns, as practitioners want transparent systems that not only perform well but also justify their outputs [139,140,141].

4.2.3. Integration into Existing Workflows

Another frequently cited barrier is the difficulty of embedding AI into existing dental workflows and clinical infrastructures. As Müller et al. [140] reported in their qualitative interviews, clinicians emphasized the practical challenge of using AI chairside if it requires additional software, disrupts normal patient flow, or increases appointment time. Integration into electronic health records, imaging software, and practice management systems is often lacking. Liu et al. [139] also noted that dentists feel inadequately trained to interpret or act on AI recommendations, suggesting that without appropriate education and user-friendly interfaces, adoption will remain low.

4.2.4. Cultural and Professional Resistance

Beyond technical and legal issues, cultural factors play a role. Dentists in the Müller et al. [140] study expressed skepticism about whether AI could complement, rather than replace, their expertise. This aligns with findings from Hoffman et al. [141] in allied health professions, where clinicians often perceive AI as undermining professional autonomy, or fear being “deskilled” if AI takes over interpretive tasks. Building trust requires framing AI as a supportive adjunct rather than a substitute for clinical judgment.

4.2.5. Resource and Infrastructure Limitations

Successful implementation also depends on resources such as high-quality digital infrastructure, internet access, and updated imaging devices. Liu et al. [139] highlight disparities between well-equipped urban practices and resource-limited rural or public clinics, which may widen inequities in access to AI-enhanced care. Hoffman et al. [141] similarly found that organizational readiness and investment in digital tools strongly influenced whether AI was realistically usable in clinical environments.

4.3. Clinical Implications for Practitioners

The clearest near-term applications of AI in dentistry are in diagnostic support. AI can help standardize caries and lesion detection, support periodontal screening, and triage radiographic findings, particularly in high-volume or resource-limited settings. Such systems may improve efficiency and reduce inter-operator variability, but they should be viewed as decision-support tools rather than replacements for clinical judgment. Over-reliance risks diagnostic complacency, de-skilling of practitioners, and failure to manage ambiguous or complex cases.

In orthodontics and prosthodontics, AI may improve planning efficiency and accuracy by automating cephalometric analysis or predicting prosthetic outcomes. However, integration into routine practice will depend on regulatory approval, cost-effectiveness, interoperability with existing digital workflows, and clinician acceptance. AI will also need to demonstrate tangible benefits for patient outcomes—not just technical accuracy—to justify adoption.

For practicing dentists considering AI integration, several evidence-based recommendations emerge:

Diagnostic Applications: Current AI systems show promise for supporting diagnostic decision-making, particularly in caries detection and periodontal assessment. However, these tools should be viewed strictly as diagnostic aids rather than replacements for clinical examination and professional judgment [139,140,141].

Treatment Planning: AI-assisted treatment planning shows potential for standardizing care and reducing planning time, but most systems require further clinical validation before routine adoption can be recommended. Orthodontic applications appear most mature for clinical implementation [139,140,141].

Economic Considerations: The cost-effectiveness of AI implementation varies significantly based on practice characteristics. Practitioners should carefully evaluate vendor claims against published evidence, as marketing materials may not accurately reflect real-world performance limitations [139,140,141].

While imaging has been the most visible entry point for AI in dentistry, its value extends far beyond radiographic interpretation. Dentistry is facing workforce shortages, increasing case complexity, and a rising demand for personalized, patient-centered care [139,140,141]. AI can help mitigate these challenges by supporting routine administrative tasks, optimizing resource allocation, and integrating diverse patient data to guide individualized prevention and treatment strategies. In this way, AI functions not only as a diagnostic tool but also as a system-level solution to improve efficiency, expand access, and tailor care to patient needs in an evolving healthcare landscape [139,140,141,142].

Although this review set out to evaluate both applications and clinical implementation of AI in dentistry, it is important to acknowledge that the majority of the included studies were diagnostic accuracy studies conducted in retrospective or laboratory-controlled settings. True clinical implementation—defined as prospective validation in diverse real-world patient populations and integration into routine workflows—remains limited. Therefore, while the findings of this review highlight the significant potential of AI to enhance diagnostic and planning tasks, they primarily represent the promise of clinical translation rather than evidence of established implementation.

4.4. Future Research Priorities

Several critical research gaps must be addressed to advance AI implementation in dentistry:

• Prospective Clinical Trials: Large-scale randomized controlled trials comparing AI-assisted versus conventional diagnosis and treatment planning are urgently needed to establish clinical benefit and cost-effectiveness.
• Diverse Population Studies: Training and validation datasets must include diverse patient populations across different demographic groups, geographic regions, and clinical settings to ensure equitable performance.
• Implementation Science Research: Studies examining integration with existing workflows, practitioner acceptance, patient outcomes, and economic impact are essential for successful clinical translation.
• Longitudinal Outcome Studies: Long-term follow-up studies are needed to assess whether AI-assisted decisions lead to improved patient outcomes compared to conventional approaches.

4.5. Computer Vision Advances Relevant to Dentistry

Recent developments in computer vision outside dentistry offer directly transferable strategies for dental imaging pipelines, which often face challenges such as small datasets, device heterogeneity, and limited annotations.

4.5.1. Few-Shot Detection and Generalization

Li et al. introduced ERNet, a multiview attention network with random-interpolation resize augmentation designed for few-shot surface defect detection [155]. By applying parallel channel attentions at different receptive fields and leveraging augmentation to preserve small or weak targets, ERNet improved generalization under limited data. Similar principles could strengthen dental AI systems in tasks such as early caries detection or identification of subtle periapical radiolucencies, where datasets are modest and diagnostic signals are easily lost during down-sampling [155].

4.5.2. Domain Shift and Imaging Heterogeneity

Wang et al. proposed CSC-SCL, a content-style control network with style contrastive learning for underwater image enhancement [156]. This approach disentangles content from style to produce domain-invariant features while preserving perceptual fidelity. Dental imaging faces analogous challenges, with variability across radiograph sensors, CBCT vendors, and exposure protocols [156]. Applying style-aware normalization and contrastive strategies could help mitigate these scanner-related differences and improve the robustness of dental AI models [156].

4.5.3. Weakly Supervised Segmentation

Apedo and Tao developed a weakly supervised segmentation framework for pavement cracks, combining adversarial image synthesis with a transformer-based encoder to capture long-range context [157]. This method reduced reliance on noisy pseudo-labels and achieved competitive pixel-wise performance with limited annotation [157]. Dental applications such as tooth and lesion segmentation, periodontal bone-loss mapping, or cephalometric landmarking could similarly benefit from generator-assisted supervision and transformer-based feature aggregation, reducing the annotation burden in dental datasets [157].

Although these innovations were developed in other domains, they directly address technical bottlenecks faced in dentistry: few training samples, domain shift between devices, and high annotation costs. Incorporating such strategies into dental AI research could accelerate the development of robust, generalizable, and clinically useful systems [155,156,157].

5. Limitations

Despite encouraging technical performance, translation into real-world practice remains limited. Most included studies were retrospective, using curated datasets with high image quality, which does not reflect the variability of daily practice. This contributes to spectrum bias, where algorithms trained on narrow datasets perform poorly in more diverse settings. This review has several limitations that should be considered when interpreting results. The heterogeneity of AI applications and outcome measures precluded quantitative meta-analysis, limiting our ability to provide precise effect estimates. Publication bias may favor positive results, and the rapidly evolving nature of AI technology means that newer developments may not be adequately captured in published literature.

The search was limited to English-language publications, potentially missing relevant research published in other languages. Additionally, most included studies were conducted in controlled research environments with carefully curated datasets, which may not reflect the challenges and variability encountered in routine clinical practice.

A key limitation of the current evidence base is the lack of standardized outcome reporting across studies. Performance measures were variably expressed as accuracy, sensitivity, specificity, or area under the curve (AUC), and no unified primary endpoint was consistently applied. This heterogeneity prevents direct comparison of effect sizes and precludes formal exploration of between-study variability. As such, the ranges reported in this review should be interpreted with caution. Future studies would benefit from adopting standardized reporting frameworks, such as STARD-AI and CONSORT-AI, to ensure consistency, reproducibility, and comparability across investigations.

6. Conclusions

AI applications in dentistry demonstrate significant promise across diagnostic, treatment planning, and predictive domains. Current evidence suggests that AI systems can achieve diagnostic accuracy comparable to or exceeding that of human practitioners in specific applications, particularly image-based diagnosis of caries and periodontal disease under controlled conditions.

However, substantial gaps remain between research applications and clinical implementation that must be addressed before widespread adoption can be recommended. Most current systems require further validation in real-world clinical settings with diverse patient populations and varying practice conditions. Critical questions about cost-effectiveness, regulatory approval, and optimal workflow integration remain largely unresolved.

For dental practitioners, AI represents an emerging technology that may enhance clinical practice rather than replace clinical expertise. As the technology matures and regulatory frameworks develop, selective adoption of well-validated AI applications may provide opportunities to improve diagnostic accuracy and treatment planning efficiency in specific circumstances.

The evidence suggests that AI will likely play an increasingly important role in dental practice over the coming decade. However, successful implementation will require careful consideration of current limitations, appropriate practitioner training, and thoughtful integration with existing clinical workflows. Most importantly, continued research focusing on prospective clinical validation, diverse population testing, and implementation science will be crucial for realizing the potential benefits of AI while ensuring patient safety and care quality.

It should be emphasized that, despite the breadth of AI applications identified, the current evidence base reflects mainly preclinical and retrospective investigations. As such, the scope of this review lies more in documenting existing applications and highlighting pathways toward clinical translation, rather than confirming widespread clinical implementation. Future research should bridge this gap through prospective validation and real-world deployment studies.

Practitioners considering AI adoption should maintain realistic expectations about current capabilities while staying informed about emerging evidence. The goal should be thoughtful integration of validated AI tools that complement rather than replace clinical judgment and expertise. This systematic review demonstrates that AI in dentistry has moved beyond experimental novelty and is showing tangible promise across multiple domains. While caries detection remains the most extensively validated area, applications in endodontics, periodontology, implantology, orthodontics, and restorative dentistry are rapidly expanding. Nonetheless, translation into routine practice demands rigorous external validation, ethical safeguards, and clinician engagement. AI should be seen not as a substitute for clinical expertise but as a powerful adjunct capable of augmenting diagnostic precision, efficiency, and accessibility in modern dentistry.