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