Toward Precision Diagnosis of Maxillofacial Pathologies by Artificial Intelligence Algorithms: A Systematic Review

Abstract

Purpose

This review highlights the potential of artificial intelligence algorithms, including machine learning (ML) and deep learning (DL), in improving the diagnosis and management of oral and maxillofacial diseases through advanced imaging techniques such as computerized tomography (CT) and cone-beam computed tomography (CBCT).

Methods

The current review was conducted on the basis of ISI Web of Science, PubMed, Scopus, and Google Scholar (2010–2024) using keywords related to radiography, MRI, CT, CBCT, ML, DL, and maxillofacial pathology, with a focus on clinical applications.

Results

The DL algorithms for detecting vertical root fractures achieved a diagnostic accuracy of 89.0% for premolars, with a sensitivity of 84.0% and specificity of 94.0%. It demonstrated an accuracy of 93% and a specificity of 88% in evaluating CBCT images. The GoogLeNet Inception v3 architecture achieved an AUC of 0.914, sensitivity of 96.1%, and specificity of 77.1% for CBCT, outperforming the panoramic radiograph, which had an AUC of 0.847, sensitivity of 88.2%, and specificity of 77.0%. CBCT demonstrated higher diagnostic accuracy (91.4%) than panoramic images (84.6%), with odontogenic cystic lesions exhibiting the highest accuracy. The U-Net-based DL algorithm achieves recall, precision, and F1 scores of 0.742, 0.942, and 0.831 for metastatic lymph nodes, and 0.782, 0.990, and 0.874 for nonmetastatic lymph nodes, respectively.

Conclusion

This study highlights the superior anatomical detail of CBCT, making it more reliable for diagnosing oral and dentomaxillofacial disorders. DL algorithms demonstrate high accuracy and sensitivity in diagnosing dental and odontogenic disorders and often outperform radiologists.

Introduction

Oral and maxillofacial pathology involves diverse diseases affecting the oral cavity, dental elements, jaws, and facial structures. Accurate diagnosis remains challenging due to the complexity of these conditions and the variability in interpretations among pathologists and radiologists. Effective management requires a comprehensive understanding of their etiology, clinical presentation, and histopathological features [1]. Common diagnostic methods in dentistry include clinical examinations to assess anatomy, physiology, and potential pathologies. These methods involve visual inspection, palpation, periodontal probing, two-dimensional (2D) radiographs (panoramic and intraoral X-rays), three-dimensional (3D) cone-beam computed tomography (CBCT), and ultrasonography [2–4]. Radiography and CBCT primarily assess hard tissues, including bone, enamel, cementum, dentin, and calculus, while offering limited information on periodontal soft tissues. They are also used to evaluate bone loss patterns and crown contours [5]. Two-dimensional imaging techniques have limitations, including higher radiation exposure, increased costs, low soft tissue contrast, susceptibility to artifacts from metallic objects, limited dynamic range, and nonquantitative data output [3]. Advanced imaging modalities, such as radiography, CT scan, CBCT, and magnetic resonance imaging (MRI), have become essential for diagnosing, differentiating, and managing maxillofacial, oral diseases [1, 6, 7] and other CNS disorders [8–11] (Fig. 1). Litjens et al. emphasized the critical role of radiologist-pathologist collaboration in reviewing extensive radiological and histological data for accurate diagnoses. This process often involves advanced techniques like immunohistochemistry, specialized staining, and positron emission tomography (PET) imaging, which are resource-intensive and demanding [12]. The increasing adoption of CT scans and CBCT in maxillofacial imaging has enhanced the application of deep learning (DL) and radiomics, facilitating early diagnosis, accurate prognosis, and effective treatment planning [13]. DL algorithms for CT/CBCT imaging have significantly improved the automated detection, segmentation, and classification of conditions such as jaw cysts, tumors, lymph node metastases, salivary gland disorders, temporomandibular joint (TMJ) pathologies, maxillofacial deformities, sinus diseases, and mandibular trauma. Many of these algorithms exhibit high accuracy and often outperform human capabilities, offering promising clinical tools for the early, precise, and personalized management of maxillofacial diseases [14] (Fig. 2). To facilitate routine clinical adoption, it is essential to address challenges such as the generalizability, explainability, and reproducibility of DL algorithms. Ongoing advancements in artificial intelligence, imaging technologies, tissue engineering, scaffold design, and microscale tools are revolutionizing diagnostic and therapeutic practices in dentistry and medicine [15]. Artificial intelligence (AI), introduced in the 1950s, aims to develop machines that emulate human behavior for complex tasks. It remains a transformative technology with significant potential in research and clinical applications [16]. AI significantly improves the accuracy and efficiency of pathology diagnoses by employing computational techniques that allow machines to learn from extensive datasets and forecast disease progression and treatment outcomes [17–19]. Machine learning (ML) employs advanced statistical techniques to detect patterns in data, allowing machines to achieve human-like intelligence through data-driven decision-making instead of explicit programming. ML algorithms adapt to evolving data, enhancing predictive performance over time. In medical imaging, ML improves diagnostic accuracy, and image quality, reduces acquisition times, and streamlines workflow and communication [20–22] (Fig. 3). ML shows promise in analyzing complex medical images like CT and MRI, as well as in predicting outcomes based on clinical and demographic data [22, 23]. AI models, including ML, DL, and computer vision, analyze medical images, patient data, and other relevant information. The global AI healthcare market is projected to reach $19.25 billion by 2026. AI diagnostic models have shown high accuracy and, in some cases, outperform human specialists [24]. CT scans are essential for evaluating maxillofacial pathology, with advanced reconstruction tools like volumetric reconstruction (VRT), maximum intensity projection (MIP), multi-planar reconstruction (MPR), and shaded surface display (SSD) enhancing visualization. MIP helps detect high-density structures, improving understanding of complex morphologies. MPR enables radiologists to track body cavities and channels, particularly in maxillofacial and dental regions. Surface rendering aids in visualizing vascular abnormalities. Modern algorithms reduce radiation dose while enhancing image quality [25, 26]. This review assesses recent evidence on the application of intelligent algorithms, including ML, DL, and neural networks, in analyzing dental radiographs, CT, MRI, and CBCT images for diagnosing and differentiating oral and maxillofacial diseases.

Fig. 1.

The history of medical imaging techniques in dentistry (oral and maxillofacial pathology). DL algorithms, which enhance the approach of signal and image processing, represent a significant advancement in artificial intelligence. Is it time to transition from natural to artificial intelligence?

Fig. 2.

The application of artificial intelligence models and algorithms in the diagnosis and management of oral and maxillofacial diseases

Fig. 3.

Comparison of imaging techniques for visualizing TMJ details. A Anterior (left) and lateral (right) radiographs of the mandible and TMJ. B Sagittal (left) and coronal (right) MRI images (1.5 Tesla scanner; GE Company, USA) of the TMJ. C A 16-slice CT scan (Siemens, Germany) of the TMJ, showing a mandibular condyle fracture. Among these techniques, MRI is the most suitable for segmenting and grading TMJ disorders with the aid of ML and DL

Methods

This review was conducted following Arksey and O’Malley’s five-step methodological framework and reported in accordance with the PRISMA-ScR guidelines. This systematic review searched databases such as ISI Web of Science, PubMed, Scopus, and Google Scholar from 2010 to 2024 using the following terms: ((Radiography[MeSH]) OR (Radiology) OR (MRI) OR (CT) OR (Panoramic) OR (OPG) OR (Periapical) OR (CBCT)) AND ((Diagnosis) OR (Maxillofacial) OR (Maxillofacial pathology) OR (Mandible)) AND ((Machine learning) OR (Image processing) OR (Deep learning) OR (Reconstruction algorithms)) AND ((Hard tissue) OR (Soft tissue) OR (Dental implant) OR (Treatment) OR (Tooth) OR (Gingiva) OR (Temporomandibular joint)) NOT (Animals[MeSH]).

Screening, Selection, and Data Extraction

The screening and selection process was conducted in multiple stages to ensure the inclusion of relevant studies. Initially, independent reviewers assessed the suitability of identified sources by screening titles and abstracts using predefined eligibility criteria. In the second stage, full-text articles of potentially relevant studies were obtained for further evaluation, with disagreements resolved by consensus. Data extracted from eligible studies included the first author’s name, publication year, study country, sample size, AI technique, study design, anatomical structure, imaging modality, validation method, and performance metrics such as the dice similarity coefficient (DSC) score. Two examiners conducted article screening, achieving substantial agreement (κ = 0.86). The selection process followed four stages: (1) exclusion of irrelevant publications and citations, (2) title and abstract screening by one reviewer, with uncertain cases cross-checked by the second examiner, (3) removal of incomplete and low-quality studies, and (4) final assessment against PIRD criteria, eliminating articles that did not meet the review’s objectives. For data extraction, the first reviewer compiled the required information, which was then re-evaluated by the second reviewer. Data were systematically collected under categories including author, year, diagnosis, sample size, test specimen, comparison parameters, AI model used, reference test, outcome measures, results, and conclusions. Extracted data were tabulated for analysis.

Selection Criteria

The inclusion criteria for study selection were as follows:

1. Studies involving medical imaging in dentistry conducted on adult individuals (over 18 years of age).

2. Studies focused on the use of AI to assist in the diagnosis or treatment of oral and maxillofacial pathology.

3. Studies examining the accuracy, effectiveness, or clinical application of AI in oral and maxillofacial pathology.

4. Studies related to the development or testing of new AI algorithms or models for use in oral and maxillofacial pathology.

5. Studies published in English.

6. Studies published between 2010 and 2024.

Results

The most commonly used ML and DL algorithms in dentistry include convolutional neural networks (CNNs), recurrent neural networks (RNNs), support vector machines (SVMs), random forest (RF), and K-nearest neighbors (K-NN) (Table 1).

Table 1.

Applied intelligent algorithms in dentistry with strengths and limitations

ML algorithm	Applications in dentistry	Strengths	Limitations
Convolutional neural networks (CNNs)	Dental radiograph analysis Tooth segmentation Caries, fractures, and pathology detection	High accuracy in image-based tasks Automated feature extraction	Requires large labeled datasets Computationally expensive
Recurrent neural networks (RNNs)	Predicting patient outcomes Behavioral pattern analysis in orthodontics	Excellent for sequential data Captures temporal relationships	Struggles with long sequences Susceptible to vanishing gradient issues
Transformer models	Natural language processing (e.g., summarizing dental records) Time-series data in patient monitoring	Handles long dependencies well Scalable to various datasets	Requires extensive computational resources
Support vector machines (SVMs)	Classification of dental diseases Identifying periodontal conditions	Effective in high-dimensional spaces Works well with smaller datasets	Sensitive to parameter selection Performance depends on kernel choice
Random forest (RF)	Prediction of treatment outcomes Analysis of risk factors in dental health	Handles mixed data types Robust to overfitting	Less interpretable compared to simpler models
U-Net	Dental X-ray image segmentation Identifying dental pulp and canal morphology	Effective for biomedical image segmentation Handles small datasets well	Limited in handling large-scale datasets High memory usage
Gradient boosting (e.g., XGBoost)	Diagnosis of dental pathologies Predicting implant success	High prediction accuracy Works well with structured data	Computationally intensive for large datasets
Long short-term memory (LSTM)	Predicting treatment responses over time Analyzing patient records for personalized care	Mitigates vanishing gradient issues Effective for temporal sequences	Resource-heavy for long sequences May require extensive tuning
DenseNet (dense networks)	Caries detection Analysis of orthodontic outcomes	Efficient parameter usage Reduces overfitting with feature reuse	Memory intensive Can struggle with very complex datasets
Generative adversarial networks (GANs)	Generating synthetic dental radiographs Augmenting training datasets	Effective in creating high-quality synthetic data	Training can be unstable Requires careful tuning
Capsule networks (CapsNets)	Analysis of dental CBCT scans Identification of root canal structures	Preserves spatial hierarchies Robust to image distortions	Computationally expensive Difficult to train
K-Nearest neighbors (K-NN)	Patient classification Predicting dental disease trends	Simple to implement Requires no training phase	Inefficient with large datasets Sensitive to noisy data
Autoencoders	Image denoising in radiographs Feature extraction for complex dental data	Reduces dimensionality effectively Useful for unsupervised learning	Risk of losing important information Requires fine-tuning
Reinforcement learning (RL)	Optimization of treatment planning Robotic-assisted surgeries	Learns from interactions Adaptive decision-making	Requires extensive training environments May not generalize well in unseen scenarios
Deep belief networks (DBNs)	Classifying dental materials Predicting material failure	Captures hierarchical features Useful for unsupervised learning	Training is computationally expensive Prone to overfitting in small datasets
Bayesian networks	Risk analysis for periodontal diseases Probabilistic reasoning in diagnosis	Incorporates uncertainty Interpretable results	Requires domain knowledge for prior probabilities Limited scalability
Natural language processing (NLP) models (e.g., BERT)	Processing patient narratives Extracting information from dental records	Excels in text analysis tasks Pretrained models available for fine-tuning	Requires significant data preprocessing Computationally intensive

A comprehensive review of 155 studies investigating the application of AI in the diagnosis of maxillofacial abnormalities was observed. The review assessed key metrics such as sensitivity, specificity, accuracy, and sample size, alongside various characteristics pertinent to oral and maxillofacial diseases, as presented in Table 2. By leveraging advanced image reconstruction algorithms and DL algorithms, AI substantially enhances diagnostic precision by addressing these artifacts. Setzer et al. indicated that DL algorithms achieved an accuracy rate of 93% and a specificity rate of 88% in the evaluation of CBCT images [9]. Fukuda et al. found that CBCT is significantly more effective than panoramic and intraoral radiographs for detecting vertical root fractures (VRFs), which predominantly occur in mandibular molars (54.8%) and mandibular premolars (17.6%) [27]. In the study by Lee et al., the DL algorithm for detecting VRFs achieved a diagnostic accuracy of 89.0% for premolars, with a sensitivity of 84.0%, specificity of 94.0%, positive predictive value (PPV) of 93.3%, and negative predictive value (NPV) of 85.5%. For molars, the diagnostic accuracy was 88.0%, sensitivity was 92.3%, specificity was 84.0%, PPV was 85.2%, and NPV was 91.3%. When both premolars and molars were combined, the diagnostic accuracy was 82.0%, sensitivity was 81.0%, specificity was 83.0%, PPV was 82.7%, and NPV was 81.4%. ROC curve analysis revealed the highest area under the curve (AUC) for the premolar model at 0.917, followed by the molar model at 0.890 and the combined model at 0.845 [28]. A study by Cui et al. (2022) collected 4938 CBCT scans from multiple hospitals and clinics in China, forming internal and external datasets to evaluate the robustness of an AI-based segmentation system. The AI system achieved high segmentation accuracy on the internal testing set, with Dice scores of 94.1% (teeth) and 94.5% (bones) and average surface distance (ASD) errors of 0.17 mm (teeth) and 0.33 mm (bones). Results on the external set were slightly lower, with Dice scores of 92.54% (teeth) and 93.8% (bones), demonstrating robustness across unseen data. Performance remained consistent across most teeth, though slightly reduced for third molars due to variability and absence in some patients. Representative segmentation results highlighted the system’s ability to adapt to diverse clinical scenarios, crucial for real-world digital dentistry. Overall, the AI system shows promise in accurately segmenting teeth and bones for various dental abnormalities in heterogeneous clinical environments [29] (Table 2).

Table 2.

Diagnostic efficacy of intelligent algorithms and models to visualization of dental and maxillofacial pathologies on radiological images

Study	Location	Sample size (patients/images)	Types of dental/oral pathology	Data acquired by imaging modality	Types of algorithms/model	Sensitivity	Specificity	Accuracy/AUC	Conclusion and outcome
Fukuda et al. [27]	Japan	330 panoramic images	Vertical root fracture (VRF)	Panoramic radiography	Convolutional neural network (CNN)-DetectNet	0.75	NA		The CNN model’s performance can potentially represent VRF damage in panoramic radiographic images, although this output was only from a single hospital image
Setzer et al. [9]	USA	20 CBCT images	Periapical lesions	CBCT images	DL Segmentation-U-Net architecture	0.93	0.88		AI can automatically differentiate lesions accurately and reduce interpretation compared to conventional algorithms
Lee et al. [28]	South Korea	3000	Dental caries	Periapical radiographic images	CNN-GoogLeNet Inception	0.81	0.83		Convolutional neural networks (CNNs), particularly when applied through supervised ML, are an effective method for diagnosing dental caries. This approach not only reduces the cost of oral health management but also increases the likelihood of preserving natural teeth
Cui et al. [29]	China	3172	Tooth and alveolar bone segmentation	CBCT images	ToothNet, MWTNet, and CGDNet/ V-Net network architecture	Tooth: 0.92 Alveolar bone: 0.93	NA	Tooth: 0.91 Alveolar bone: 0.93	Based on a large-scale dataset from multicenter clinics, the DL-based AI system is a robust tool for fully automated tooth and alveolar bone segmentation. It also reduces the need for manual annotation and inspection of radiological images by dentists
Choi et al. [32]	South Korea	571	Estimate the exact position between the mandibular third molar (M3) and the inferior alveolar nerve (IAN)	Panoramic radiography and CBCT	CNN (ResNet-50 architecture)	True contact position between M3 and IAN: 0.85 bucco-lingual position between M3 and IAN: 0.87	True contact position between M3 and IAN: 0.55 bucco-lingual position between M3 and IAN: 0.75	True contact position between M3 and IAN: 0.72 bucco-lingual position between M3 and IAN: 0.81	The DL algorithm accurately determines both positions better than oral and maxillofacial surgery (OMFS) experts, presents the surgical plan correctly, and minimizes CBCT radiation exposure in line with the ALARA principle
Poedjiastoeti W, Suebnukarn S. [38]	Thailand	500	Odontogenic tumors of the jaw	Panoramic radiography	CNN (VGG16)	0.82	0.83	0.83	CNN training algorithms could screen ameloblastomas and keratocystic odontogenic tumors with high accuracy comparable to oral maxillofacial specialists in a substantially shorter time
Lee et al. [8]	South Korea	912 panoramic radiography and 789 CBCT images	Odontogenic cystic lesions (OCLs)	Panoramic and CBCT images	CNN (GoogLeNet Inception v3 architecture)	Panoramic: 0.88 CBCT: 0.96	Panoramic: 0.77 CBCT: 0.77	Panoramic: 0.85 CBCT: 0.91	The AI system effectively detects and diagnoses OCLs using panoramic radiography and CBCT image datasets. However, the diagnostic accuracy of OCLs based solely on radiological assessment is lower than that achieved with histological examination, and accurate diagnosis using radiological images alone remains challenging. Additionally, the GoogLeNet Inception v3 architecture demonstrated significantly higher diagnostic efficacy on CBCT images compared to panoramic images
Chai et al. [39]	China	272	Ameloblastoma and odontogenic keratocyst	CBCT images	CNN (Inception v3 DL algorithm)	0.87	0.82	0.85	The DL algorithm demonstrates high performance in differentiating ameloblastoma (AME) and odontogenic keratocyst (OKC) compared to oral and maxillofacial specialists. Additionally, the AI system can significantly assist in noninvasive surgical approaches and therapeutic planning, providing results more quickly than both senior and junior surgeons
Ariji et al. [40]	Japan	703 images from 51 patients	Lymph node metastases in patients with oral squamous cell carcinoma	Contrast-enhanced CT	CNN (AlexNet)	0.67	0.90	0.84	The AlexNet DL algorithm has low inter-model variability bias compared to radiologists’ interobserver variation, so higher reliability and significant diagnostic performance than that of radiologists
Ariji et al. [30]	Japan	672 images from 51 patients	Metastatic cervical lymph nodes in patients with oral cancers	Contrast-enhanced CT	CNN (U-Net)	0.98	0.95	0.96	Although the U-Net DL model is in its infancy in terms of segmentation performance and needs improvement, it was able to accurately segment lymph node metastases compared to radiologists
Kise et al. [35]	Japan	400 images (200 from 20 Sjögren’s syndrome patients and 200 from 20 control subjects)	Sjögren’s syndrome	CT scan images	CNN (AlexNet)	100	0.92	0.96	AlexNet, as a DL algorithm, exhibits diagnostic performance comparable to experienced radiologists, making it a valuable tool for interpreting CT images and a reliable consultant in therapy teams
Zhang et al. [41]	China	1320 CT images from 132 patients	Classification of benign and malignant parotid tumors	CT scan images	Improved CNN model /VGG16, Inception v3, ResNet, DenseNet	0.97	0.99	0.98	The improved CNN model, compared to four classic pretraining models including VGG16, Inception v3, ResNet, and DenseNet, effectively diagnoses benign and malignant parotid tumors. It enhances surgeons’ decision-making capabilities regarding intervention methods and prognosis monitoring
Wang et al. [42]	China	408 images from 686 patients	Detection of mandibular fractures in nine subregions	Panoramic radiography and CT scan images	CNNs (U-Net and ResNet)	0.91–0.97	0.91–0.99	0.94–0.98	U-Net and ResNet models are reliable and accurate tools for detecting and classifying mandibular fractures. It is suggested that DL models especially U-Net due to the high DICE (0.94) in automatic segmentation practically used in dentistry clinics lacking experienced doctors
Ezhov et al. [43]	Cyprus	1346 scans	Detection of caries, periapical lesions, and periodontitis	CBCT	CNN U-net-like architecture(Diagnocat commercially platform)	0.92	0.99	NA	The evaluation of large-scale imaging data by Diagnocat has been improved, reducing the time required for interpretation and differential diagnosis. AI-based platforms, such as Dentaverse, have the potential to be integrated into routine dental practice. However, it is important to note that these DL algorithms must be approved by scientific regulatory organizations before they can be widely used in clinical settings
Kubo et al. [44]	Japan	161 patients	Prediction of metastasis in patients with cervical lymph node cancer	Contrast-enhanced CT	Key ML algorithms including K-nearest neighbor (K-NN), SVM, CART, RF, and AdaBoost	0.82	NA	0.85	The radiomics-based predictive model, utilizing ML, exhibited outstanding diagnostic performance in tracking cervical lymph node metastasis. In particular, the support vector machine (SVM) model achieved an impressive AUC score of 0.98, highlighting its potential as a valuable clinical decision-support tool
Bianchi et al. [45]	USA	92 patients	Diagnosis of TMJ osteoarthritis	CBCT images	ML approaches (extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), random forest)	0.84	NA	0.82	The two ML algorithms, LightGBM and XGBoost, demonstrate strong diagnostic performance and reliability in predicting the status of temporomandibular joint (TMJ) osteoarthritis. By integrating clinical, imaging radiomics, and patient-specific biomolecular data, a targeted diagnosis and personalized treatment plan can be effectively developed
Mori et al. [33]	Japan	350	Evaluation of the positioning in periapical radiography of the maxillary teeth	Periapical radiography	AlexNet and U-net	0.92	0.82	0.87	In cases where the root apex of the maxillary canine could not be properly segmented in periapical radiographs, the DL algorithm detected the image as poor quality. This helped prevent misinterpretation by maxillofacial radiologists and reduced the risk of inappropriate treatment
Esmaeilyfard et al. [46]	Iran	819 CBCT images from 150 patients	Periapical cysts and dental caries	CBCT images	CNN (Adam algorithm)	0.90	0.99	0.98	DL algorithms as an important strategy can accurately detect dental caries on CBCT images, so may improve dentist diagnostic outcomes and treatment planning in clinics with lower costs
Alalharith et al. [47]	Saudi Arabia	134 intraoral images	Dental plaque	Dental photography	ResNet-50 CNN/Faster R-CNN model	–	–	100%	DL algorithms play a crucial role in diagnosing gingivitis in intraoral images, contributing to the global reduction of periodontal disease severity through early detection and preemptive diagnosis
Lee et al. [48]	Korea	1044	Periodontal disease	Periapical radiography	VGG-19 network/deep CNN model	–	–	Molar: 0.73 Premolar: 0.83	The VGG-19 network is an efficient method for diagnosing and predicting periodontally compromised teeth
Krois et al. [49]	Germany	1809	Alveolar bone loss	Panoramic radiograph	CNN/deep feed-forward	0.81	0.81	–	The CNN algorithm, when compared to six experienced dentists, demonstrated equivalent diagnostic efficacy in assessing periodontal bone loss (PBL) on panoramic radiographs
Kim et al. [50]	South Korea	12,179 radiographs	Alveolar bone loss	Panoramic radiograph	CNN (DeNTNet)	0.78	0.92	–	The DeNTIsts algorithm achieves superior periodontal bone loss (PBL) detection performance compared to dental clinicians. It can reduce dentist's workload and improve the interpretation of large volumes of radiographic images
Kühnisch et al. [51]	Germany	2,417 images	Dental caries	Intraoral images	CNNs	0.89	0.94	0.92	This study concludes that, under ideal conditions, caries detection using intraoral photographs with a trained CNN can achieve over 90% agreement. Additionally, other pathological findings, such as developmental defects or dental restorations, must be considered when using clinical photographs as a machine-readable diagnostic tool
Duong et al. [52]	Vietnam	620 unrestored molars and premolars teeth	Dental caries	Smartphone images	SVM	0.88	0.96	0.92	In this study, we developed an automated methodology to detect and classify dental caries lesions using smartphone color images based on the ICDAS II system. The results led to the creation of an AI-integrated mobile application, which can provide early warnings about oral health conditions, enabling timely treatment. Additionally, the app can assist dentists by reducing clinical examination time and providing more accurate, unbiased diagnoses
Zheng et al. [53]	China	844 radiographs	Depth of carious lesions related to pulpitis	Periapical radiographs	CNN of ResNet18	0.89	0.86	0.86	The ResNet18 CNN demonstrated strong performance in diagnosing deep caries and pulpitis. The multimodal ResNet18 CNN further enhanced performance, showing promising potential for diagnosing these conditions
Chauhan et al. [54]	India	428 radiographs	Depth of carious lesions related to pulpitis	Periapical radiographs	CNN-fuzzy-based	–	–	0.94	The computer-aided decision-making system for pulpitis enhances dentists’ confidence in diagnosing reversible and irreversible pulpitis, while minimizing false diagnoses caused by ambiguous radiographic values, signs, and symptoms
Maharani et al. [55]	South Korea	21 periapical radiographs	GLCM* and watershed image segmentation methods	Periapical radiographs	K-NN	–	–	0.83	Teeth were classified using K-nearest neighbors (K-NN) to categorize images into three groups: reversible pulpitis, irreversible pulpitis, and normal teeth. The results indicate that the optimal dimension for extracting periapical radiographs is 256 × 256, which yields the highest accuracy compared to other dimensions. During the classification phase, it was found that a lower K value resulted in higher accuracy
Kirnbauer et al. [36]	Austria	144 CBCT images	Periapical lesions	CBCT	U-Net	0.97	0.88	–	Although periapical lesions (PALs) exhibited variations in appearance, size, and shape in the CBCT dataset, the U-Net algorithm as an automated method demonstrated excellent results compared to routine protocol
Yuce et al. [56]	Turkey	2000 bite-wing radiographs	Pulpal calcification	Bite-wing radiographs	YOLOv4	–	0.97	0.96	The YOLOv4 algorithm, trained on bite-wing radiographs, successfully detected pulp chambers and calcifications with high accuracy

The study by Lee et al. evaluated the performance of a deep convolutional neural network (CNN) based on the GoogLeNet Inception v3 architecture for detecting and diagnosing three odontogenic cystic lesions (OCLs): odontogenic keratocysts, dentigerous cysts, and periapical cysts. The dataset included 2,126 images (53.6% panoramic, 46.4% CBCT), with preprocessing steps like resizing, normalization, and augmentation to enhance model robustness. The CNN achieved an AUC of 0.914, sensitivity of 96.1%, and specificity of 77.1% for CBCT images, outperforming panoramic images, which had an AUC of 0.847, sensitivity of 88.2%, and specificity of 77.0%. CBCT demonstrated higher diagnostic accuracy (91.4%) compared to panoramic images (84.6%), with periapical cysts showing the highest accuracy. The study highlights CBCT’s superior anatomical detail, making it more reliable for detecting OCLs, though limitations included a small dataset and the exclusion of ameloblastoma cases. These findings support the potential of DL for automated dental diagnostics [8].

Ariji et al. developed a U-Net-based DL algorithm for automatic segmentation of cervical lymph nodes and diagnosis of metastases in oral squamous cell carcinoma patients using contrast-enhanced CT images. The dataset included 158 metastatic and 514 nonmetastatic lymph nodes from 59 patients, divided into training, validation, and testing sets. Trained for 200 epochs with the Adam optimizer, the model achieved recall, precision, and F1 scores of 0.742, 0.942, and 0.831 for metastatic lymph nodes, and 0.782, 0.990, and 0.874 for nonmetastatic lymph nodes, respectively. The diagnostic accuracy for metastasis reached an AUC of 0.950, surpassing radiologists’ performance (AUC of 0.896). While the model demonstrated high precision, its recall was lower, especially for small lymph nodes and difficult cervical regions, suggesting areas for further improvement [30] (Table 2).

Discussion

AI is a highly effective tool in the management of orthodontic conditions. In orthodontics, AI is utilized for treatment planning and outcome prediction, including the simulation of changes in facial photographs taken before and after treatment. AI algorithms can also analyze the effects of orthodontic interventions, assess skeletal patterns, and identify anatomical landmarks in lateral cephalograms, thereby improving communication between clinicians and patients and enhancing the overall treatment process [31]. The ability of MIP to distinguish hyperdense structures from surrounding tissues aids in the interpreter’s understanding of the morphology of intricately superimposed structures.

Accurate and timely detection of dental caries is essential for effective prevention and treatment. Undiagnosed caries can progressively damage enamel, dentin, and pulp, leading to severe pain and tooth loss. A deep convolutional neural network (CNN) based on the GoogLeNet Inception v3 architecture demonstrated significant improvements in caries detection accuracy and efficiency, supporting the integration of optimized DL algorithms into clinical dental practice [28]. Radiography, combined with clinical examination, is vital for detecting interproximal and root caries. Deep CNN algorithms, however, excel in edge detection through multiple convolutional and hidden layers, enabling improved identification of caries location and morphology [32, 33]. Advancements in DL, such as ResNet which has deeper and more complex layers, have significantly enhanced the accuracy of maxillofacial detection and dental segmentation [34]. The U-Net-based model developed by Ariji et al. outperformed experienced radiologists, demonstrating its potential to improve clinical workflows with rapid, objective, and accurate results. The model provided diagnostic predictions within 7 s, enhancing efficiency. However, challenges remain, including difficulties in identifying smaller lymph nodes and addressing cervical anatomy variability, which affects recall performance. This study emphasizes the need for further research with multicenter datasets and multimodal imaging to improve generalizability and robustness. Despite limitations such as a single-center dataset and imaging constraints, the findings highlight the promise of DL for accurate and efficient lymph node segmentation and metastasis detection in diverse clinical settings [30]. Cui et al. developed a fully automated AI system for dental ROI localization, tooth segmentation, and alveolar bone segmentation directly from raw CBCT images without manual intervention. The system integrates tooth skeleton information for geometric guidance and employs auxiliary tasks, such as boundary and root landmark prediction, to increase accuracy in low-contrast scenarios. Designed for orthodontic applications, the model ensures that tooth root apices remain confined within alveolar bones during movement. A filter-enhanced cascaded network utilizing the Harr transform improves bone segmentation by enhancing the contrast between alveolar bones and soft tissues. When data-driven and knowledge-driven approaches are combined, the model demonstrates high accuracy and robustness, highlighting its clinical potential [29]. Lee et al. revealed the effectiveness of a deep CNN architecture on CBCT images because of enhanced anatomical detail and fewer artifacts than panoramic images. Challenges included variability in lesion shape, size, and boundary clarity, complicating automatic segmentation. Limitations, such as a small dataset and the exclusion of ameloblastoma cases, highlight the need for improvement [8].

The interactive tool MPR enables tracking of body foramina, particularly in the maxillofacial and dental regions. Surface rendering is an effective postprocessing technique for visualizing vessel anomalies. Compared with conventional CT, advanced reconstruction algorithms reduce the radiation dose while improving image quality, and CBCT offers higher resolution and lower susceptibility to metal artifacts, increasing its diagnostic value. Additionally, AI algorithms support clinical decision-making by analyzing patient data, medical histories, and imaging results to suggest treatment plans and predict outcomes, aiding dentists in diagnosing and managing oral and maxillofacial pathologies. Compared with radiologists, Kise et al. evaluated the diagnostic performance of a DL algorithm using the AlexNet architecture for detecting Sjögren’s syndrome (SjS) in CT images. The model, trained on 500 CT images (200 from 20 SjS patients and 200 from controls), achieved 96.0% accuracy, 100% sensitivity, 92.0% specificity, and an AUC of 0.960 after 300 training epochs. Experienced radiologists achieved slightly higher accuracy (98.3%) with comparable sensitivity and specificity, whereas inexperienced radiologists achieved significantly lower performance. The DL system demonstrated equivalence to experienced radiologists and outperformed inexperienced radiologists, suggesting its potential to assist clinicians, particularly those who are less experienced, in reliably diagnosing SjS. Challenges include dataset limitations, reliance on compressed image formats, and a lack of pathological variety in controls [35]. Kirnbauer et al. developed a DL framework that combines SpatialConfiguration-Net (SCN) for tooth localization and a modified U-Net for detecting periapical osteolytic lesions (PALs) in CBCT images. The SCN achieved high localization accuracy (97.3%) with a mean point-to-point error of 1.74 mm, ensuring precise tooth segmentation. The modified U-Net demonstrated strong performance in lesion detection, with sensitivity and specificity of 97.1% and 88.0%, respectively, and a Dice score of 66.5% for lesion segmentation. The model effectively addresses challenges like class imbalance using focal loss, improving the detection of small lesions. Compared with other studies, it showed superior diagnostic performance, highlighting its potential for reliable, automated PAL detection in clinical settings. However, challenges, such as low contrast and lesion variability, remain, suggesting a need for further improvements in dataset diversity and model refinement [36].

In oral and maxillofacial pathology, AI holds significant potential for improving diagnostic accuracy and enabling personalized care. However, as a relatively new technology, its clinical application requires careful consideration of ethical, legal, and regulatory challenges.

Clinical Implications of AI in Dental Practice

The integration of AI into routine dental and maxillofacial practice presents transformative opportunities for precision diagnostics, treatment planning, and workflow optimization. However, the successful clinical translation of AI models necessitates a structured implementation strategy that fosters collaboration between radiologists, dentists, and AI developers. This section addresses the clinical implications of AI in dental radiology and provides guidance on how professionals should integrate AI-driven tools into their practice.

AI has demonstrated remarkable accuracy in detecting oral and maxillofacial pathologies, often surpassing traditional radiographic interpretation by human specialists. Convolutional neural networks (CNNs) and DL models have been particularly effective in the automated detection of caries, periodontal diseases, periapical lesions, and cystic pathologies in CBCT and panoramic radiographs [8, 9]. Despite these advances, clinicians must remain engaged in AI-driven diagnostics to ensure accurate interpretation, considering that AI models are trained on pre-existing datasets that may not encompass the full spectrum of patient variations encountered in clinical practice.

To facilitate AI adoption, dental professionals should undergo training in AI-assisted image analysis and interpretation. Continuing education programs, workshops, and interdisciplinary collaborations between dental schools, AI researchers, and radiology departments can bridge the knowledge gap. Moreover, AI tools should be designed with user-friendly interfaces that allow seamless integration into digital imaging and communications in medicine (DICOM) workflows, ensuring efficiency in diagnostic procedures [31, 34].

From a clinical workflow perspective, AI models can function as adjunct tools that assist radiologists and dentists in decision-making rather than replacing human expertise. For instance, AI-based segmentation of cystic lesions or tumor margins in CBCT scans can enhance diagnostic accuracy while reducing interpretation time. Additionally, AI-powered treatment planning systems can analyze patient-specific imaging data to propose orthodontic interventions, implant placements, and surgical procedures with enhanced precision [29, 30]. However, it is imperative that AI-generated recommendations be validated by human experts to avoid potential errors arising from dataset biases or software limitations.

Ethical considerations and regulatory compliance also play a crucial role in the implementation of AI in dental practice. AI-driven diagnostic tools must adhere to international standards for medical device approval, including guidelines from the US Food and Drug Administration (FDA) and the European Medicines Agency (EMA) [35]. Furthermore, issues related to patient data privacy and informed consent must be addressed, particularly when AI models are trained on large-scale anonymized datasets.

AI has the potential to revolutionize dental diagnostics by enhancing accuracy, efficiency, and personalized treatment planning. However, its successful integration into routine practice requires a collaborative effort among dental professionals, radiologists, and AI developers. By embracing AI as an assistive tool rather than a replacement for clinical expertise, the dental community can harness its full potential while maintaining high standards of patient care.

Conclusion

Despite promising results, challenges, such as low contrast and lesion variability, persist, necessitating further model refinement. Future research should expand the dataset to include a wider range of clinical scenarios, diverse lesion types, and varying imaging conditions to improve its robustness and generalizability. These advancements could enable the integration of DL algorithms into routine dental workflows, improving diagnostic accuracy and efficiency. While these algorithms have been shown to outperform radiologists in diagnosing dental and odontogenic disorders, limitations, such as false positives and negatives, remain. Future work should focus on refining segmentation accuracy and expanding datasets to enhance model reliability. Eventually, DL algorithms have the potential to reduce radiologists’ reading time, improve diagnostic precision, and assist surgeons with intraoperative reference maps, making them valuable tools in clinical practice.

Risk of Bias Assessment

The implementation of AI in CBCT airway analysis faces key challenges, primarily due to the limited availability and comprehensiveness of training datasets. While AI research typically utilizes thousands of datasets, CBCT airway studies often rely on only 40 to 300 scans, with most using a single CBCT machine. Additionally, training and testing are typically conducted within a single center using supervised learning, limiting the generalizability of AI models. Environmental concerns, particularly high carbon emissions from AI model training, further highlight the need for sustainable alternatives. AI accuracy is highly dependent on CBCT data quality, which can be affected by patient motion, head positioning, artifacts, kV parameters, and field of view limitations. Technical factors, such as electrical noise, scatter, and beam hardening, further compromise image accuracy. Improving AI performance requires enhanced imaging protocols, diverse datasets, and clinical context integration. Bias remains a concern in AI research, with 11% of studies in this review reporting a high risk of bias in reference standards. However, AI was classified as a low-risk technology due to its reliance on standardized data inputs, with minimal bias observed in index testing and flow and timing. Tools, such as the Prediction Model Risk of Bias Assessment Tool (PROBAST), help evaluate bias and applicability in AI model studies. Additionally, guidelines, like the MI-CLAIM checklist and the Checklist for Artificial Intelligence in Medical Imaging (CLAIM), provide standardized frameworks for reporting AI model development, validation, and application in clinical settings, ultimately reducing bias in AI modeling.

Ethical Issues and Limitations in AI

This review carefully applied filters to ensure relevance without significantly impacting the retrieval of pertinent studies, despite the general discouragement of such restrictions in systematic review methodology. The inclusion criteria were limited to human studies published between 2013 and 2022, reflecting recent advancements in neural network applications in dental imaging. While this approach enhances relevance to contemporary research, it may exclude historically significant works published before 2013. AI development for CBCT airway analysis faces challenges due to the lack of standardized methodological frameworks and ethical concerns surrounding medical data usage. Patient imaging data are highly sensitive and subject to strict privacy regulations, particularly under the EU General Data Protection Regulation (GDPR), which requires explicit consent for AI research and renewal for each algorithm update. In contrast, some regions allow imaging centers to own and share data more easily for AI development. Additionally, the commercialization of AI trained on medical data raises ethical and regulatory concerns, requiring ongoing scrutiny. Environmental sustainability is another critical issue, as AI computations demand high energy consumption, contribute to greenhouse gas emissions, and require substantial water usage, is a critical concern. Addressing these challenges necessitates a balanced approach that ensures ethical AI development while considering data accessibility, regulatory compliance, and environmental impact. For example, training large language models, such as GPT-3, consumes 5.4 million liters of water for server cooling, whereas an average of 30 prompts in ChatGPT requires approximately 500 ml of water [37].

Funding

The authors did not receive any funding.

Declarations

Conflicts of interest

All authors have no conflicts of interest.

Ethics Approval and Consent to Participate

Not applicable.