Advancements in Caries Diagnostics Using Bitewing Radiography: A Systematic Review of Deep Learning Approaches

Abstract

Introduction

Deep learning techniques have emerged as promising tools for enhancing the radiographic diagnosis of caries, particularly when utilizing bitewing radiographs.

Methods

Following the PRISMA guidelines, a systematic review was conducted to assess the use of deep learning for caries diagnosis in bitewing radiographs. Literature searches were performed across Web of Science and PubMed databases for studies published before March 2025 that utilized deep learning for caries detection, segmentation, and classification using bitewing radiographs. Data extraction focused on model architectures, dataset characteristics, annotation processes, diagnostic performance metrics, and potential biases, as assessed by the QUADAS-2.

Results

Twenty-three studies met the inclusion criteria, encompassing caries detection, segmentation, and severity classification. The most frequently applied deep learning models were classification models, such as ResNet and detection models, such as YOLO architectures. Dataset sizes varied widely, ranging from 112 to 8,539 images. Most studies reported high diagnostic performance, with accuracies ranging from 70% to 99%. Some AI models outperformed or matched the performance of human experts, particularly in detecting advanced carious lesions. However, considerable variability was observed in model architectures, dataset characteristics, the applied diagnostic performance metrics, and reporting standards. The risk of bias assessment revealed concerns in patient selection, index test interpretation, and reference standards, with all studies rated as having a high risk of bias in at least one domain.

Conclusion

The review identified challenges in currently developed deep learning models regarding methodological heterogeneity, lack of standardization, limited dataset diversity, insufficient clinical validation, and concerns about bias and data transparency. Nevertheless, all studies concluded that deep learning models are promising as an assistive diagnostic tool in caries diagnostics using bitewing radiography.

Plain Language Summary

This systematic review examines the use of deep learning techniques for detecting, segmenting, and classifying dental caries in bitewing radiographs. The review included 23 studies published before March 2025, analysing the performance of various deep learning models. The most commonly used models were ResNet and YOLO, with dataset sizes ranging from 112 to 8,539 images. Most studies showed high diagnostic accuracy, ranging from 70% to over 99%, and some AI models even outperformed or matched human experts, especially in detecting advanced caries. However, there were significant differences between the models, datasets, and the applied performance metrics. Many studies had potential biases, particularly in patient selection and test interpretation. Despite these challenges, the review concluded that deep learning models hold great potential as supportive tools for caries diagnosis using bitewing radiographs, though further standardization and clinical validation are needed.

Introduction

Early detection and staging of caries are crucial for effective treatment and prevention of their progression [1]. Conventional methods for caries diagnosis include visual-tactile examinations and radiographic assessments using the bitewing technique [2]. However, radiographic diagnosis of caries presents several challenges. Variations in image quality, overlapping anatomical structures, and the subtle appearance of early-stage lesions can make accurate diagnosis difficult, even for experienced clinicians [3]. Moreover, radiographic interpretation can be time-consuming and prone to subjectivity, leading to potential inconsistencies in diagnosis [3].

To address these challenges, integrating deep learning, a recent subset of artificial intelligence, in dental imaging has emerged as a promising approach to assist clinicians and enhance the performance of radiographic caries diagnostics, ultimately improving treatment outcomes [4]. Deep learning can be applied in radiographic caries diagnostics through three key tasks: detection, segmentation, and classification. Detection focuses on identifying the presence and location of caries, with the AI typically marking the affected area using a bounding box or marker around the detected region. Segmentation takes this a step further by locating the lesion and precisely outlining its exact boundaries and shape. Classification models categorize the lesions into sub-classes by displaying textual information and color-coded labels directly on the radiograph.

In recent years, an increasing number of studies have investigated the application of deep learning in caries prediction and detection, utilizing visual inspection, clinical photographs, and radiographs [5]. To our knowledge, one recently published meta-analysis has explicitly examined the diagnostic performance of deep learning models for caries detection using bitewing radiographs, while another focused solely on their performance in diagnosing approximal caries. [6, 7]. In addition to diagnostic performance, this current study aims to review the deep learning models systematically developed for all types of caries diagnostics using bitewing radiography, considering model architectures, dataset characteristics, annotation processes, and bias assessment. Addressing this gap is crucial for understanding the current state of research, identifying its limitations, and informing future studies.

Methods

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [8], and the PRISMA 2020 checklist was followed (online suppl. Table S1; for all online suppl. material, see https://doi.org/10.1159/000546448). The search strategy was reviewed by both K.S.H.H and X.Q.S for accuracy using the Peer Review of Electronic Search Strategies (PRESS) criteria [9].

Literature Search Strategy

A comprehensive literature search was conducted to identify relevant articles on the application of deep learning algorithms for caries detection using bitewing radiographs. The databases searched included PubMed and Web of Science. The search was limited to articles published in English between January 2019 and March 2025. Initial searches were conducted in February 2023, with updates in March and August 2023, May 2024, and March 2025. All references were managed using EndNote 21 software (Clarivate Analytics). Detailed search queries and the number of results for each database are provided in online supplementary Tables S2 and S3.

Inclusion Criteria

Studies published in English between January 2019 and March 2025, including peer-reviewed original research articles, utilized deep learning algorithms for caries detection in bitewing radiographs and reported diagnostic metrics.

Exclusion Criteria

The exclusion criteria applied in case of not reporting sample size; “self-trained” models without involvement of dentists’ diagnosis; studies with fewer than 100 images; studies using other radiographic modalities (e.g., panoramic radiography, CBCT).

Study Selection

A total of 1,101 articles were obtained from the database searches. After the removal of duplicates, 425 unique records remained. These were uploaded to the Rayyan Web application for systematic reviews to streamline the review process [10]. The titles and abstracts were screened independently by two reviewers (K.S.H.H and X.Q.S) to identify studies that met the inclusion criteria. Full-text articles were obtained for the studies that appeared relevant or when eligibility was uncertain. Any discrepancies between the reviewers were resolved through discussion or consultation with a third reviewer (N.R.G). The search strategy is presented in the PRISMA flowchart (Fig. 1).

Fig. 1.

PRISMA flowchart of the article selection and inclusion process.

Data Extraction

Data were extracted independently by K.S.H.H and X.Q.S using a standardized data extraction form developed for this study. Any uncertainty during data extraction was resolved through discussion or consultation with the co-authors. The extracted data included study characteristics (authors, publication year, country); dataset characteristics (number of images, augmentation techniques, dataset allocation, and lesion prevalence in training data); deep learning models (model architecture, model task, training parameters, and training time); annotation details (number and clinical experience of annotators, annotation tasks performed); data availability (whether the dataset is publicly available or accessible upon request); diagnostic performance metrics of deep learning models and any comparisons with human performance. Dentists are familiar with diagnostic metrics such as accuracy, sensitivity, and specificity, which align well with clinical diagnostic needs. However, when developing deep learning models for medical imaging, metrics such as F1 score, recall (also known as sensitivity), precision (positive predictive value), average precision (AP), and intersection over union (IoU) are often used. Table 1 illustrates the commonly used performance metrics, their formulas, and what they measure.

Table 1.

Commonly used performance metrics, their formulas, and their corresponding interpretations in imaging diagnostics

Metric	Formula	What it measures
Accuracy	$\frac{TP+TN}{TP+TN+FP+FN}$	Proportion of correctly classified instances
F1 score	$\frac{2\text{⋅}Precision\text{⋅}Recal\mathrm{l}}{Precision+recall}$	Balance of precision and recall
Intersection over Union	$\frac{AreaofOverlap}{AreaofUnion}$	Overlap between predicted and ground truth regions
Sensitivity, recall	$\frac{TP}{TP+FN}$	Proportion of actual positives correctly identified
Specificity	$\frac{TN}{TN+FP}$	Proportion of actual negatives correctly identified
Precision positive predictive value	$\frac{TP}{TP+FP}$	Proportion of predicted positives that are actually positive
Negative predictive value	$\frac{TN}{TN+FN}$	Proportion of predicted negatives that are actually negative
False-positive rate	$\frac{FP}{FP+TN}$	Proportion of actual negatives incorrectly identified as positives
False-negative rate	$\frac{FN}{FN+TP}$	Proportion of actual positives incorrectly identified as negatives
Average precision	Area under the precision-recall curve	Model's precision at all points where recall increases
Area under the curve	Area under the receiver operating characteristic curve	Model’s ability to distinguish between classes across all classification thresholds

Data Evaluation

Descriptive Analyses on Dataset Characteristics, Deep Learning Models, Annotation Details, Reported Performance Metrics, and Risk of Bias Assessment

The quality and risk of bias of the included studies were assessed independently by K.S.H.H using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool [11]. The risk of bias was assessed in the following across four domains for each of the studies: patient selection, index test, reference standard, and flow and timing. In addition, applicability concerns were also assessed for the first three domains. Any uncertainty during bias assessment was resolved through discussion or consultation with co-authors.

Results

Twenty-eight unique records were identified for full-text review [12–39], of which 23 were deemed suitable for inclusion [12–34]. Two studies were excluded because they used periapical radiographs [36, 37]; one did not use a gold standard or validate against dentists, making it unclear how they achieved their results [38]. The other two focused on commercially available deep-learning products or Web-based services [35, 39].

The extracted data from the 23 studies are summarized in Tables 2 a, b. Table 2a provides clinical details regarding the origin of the radiographs, patient demographics, radiographic acquisition systems, and exposure parameters. Table 2b presents information on the applied deep learning architectures, including annotator characteristics (number and clinical experience), prevalence and types of caries, and diagnostic performance metrics of the AI models.

Table 2.

Summary of the included studies and their characteristics

Authors	Bitewing origin	Patient demographic	Intraoral X-ray unit	Image receptor	Exposure parameters
a Included studies on the origin of the retrieved bitewing radiographs, patient demographics, radiographic system, image receptor used, and exposure parameters
Ayhan (2024) [14]	Kırıkkale University	Age: >18	Gendex Expert DC (Gendex Dental Systems, ILB, USA)	–	65 kVp, 7 mA
	Faculty of Dentistry
	Turkey
Karakuş (2024) [26]	Necmettin Erbakan University, Faculty of Dentistry, Turkey	–	Morita Veraview X Type R Intraoral Digital Imaging Device (J. Morita Corp., Kyoto, Japan)	Phosphor plates: Soredex Digora Optime image scanner system from Finland	70 kVp, 7 mA, 0.25 s
de Frutos (2024) [23]	Six Norwegian institutes and research centres, Norway	56% female 44% male	–	–	–
		Age: 19–94
Azhari (2023) [15]	King Abdulaziz University Dental Hospital in Jeddah and dental practitioner, Saudi Arabia	–	–	–	–
Ahmed (2023) [12]	King Abdulaziz University Dental Hospital in Jeddah, Saudi Arabia	–	–	–	–
Panyarak (2022) [31]	Chiang Mai University, Faculty of Dentistry, Thailand	Permanent teeth	Heliodent Plus (Dentsply Sirona, Charlotte, NC, USA)	Carestream 7600 Phosphor Plate (Carestream, Rochester, NY, USA) and VistaScan Imaging Plates Plus (DÜRR DENTAL, Bietigheim-Bissingen, Germany)	–
Kunt (2023) [27]	A hospital information system (not specified)	Adult	–	Three direct radiography and one indirect radiography (manufacture not specified)	–
Vimalarani (2022) [34]	A faculty (not specified)	–	–	–	–
Mao (2021) [29]	–	–	–	–	–
Moran (2021) [30]	–	–	Sirona Heliodent Plus oral X-ray unit (Kavo Brasil Focus)	The EXPRESS™ Origo imaging plate (KaVo Dental, Biberach an der Riss, Germany)	70 kVp, 7 mA, 0.25–0.64 s
Bayrakdar (2022) [18]	Ordu University, School of Dentistry, Turkey	Age: >18	Kodak CS 2200 (Carestream Dental, Atlanta, GA)	2 different phosphor plate systems: Carestream CS 7600(Carestream Dental, Atlanta, GA) and Kodak CR (Carestream Dental, Atlanta, GA)	50 kVp, 5 mA, 0.1 s
Forouzeshfar (2024) [25]	Samin Oral and Maxillofacial Radiology Center in Tehran, Iran	–	Planmeca periapical device, Helsinki, Finland	–	60 kVp, 8 mA, 0.128 s
Cantu (2020) [20]	Dental clinic at Charité-Universitätsmedizin, Berlin, Germany	52% male	Orthophos XG, Dentsply Sirona (Bensheim, Germany), and Dürr Dental (Bietigheim-Bissingen, Germany)	–	–
		48% female
		Mean age: 36.5
		Age: 14–89
Chaves (2024) [21]	Dutch dental practice-based research network, The Netherlands	–	–	–	–
Chen (2022) [22]	Peking University School and Hospital of Stomatology in Beijing, China	–	Gendex Expert DC system (Gendex Dental Systems, Hatfield, PA, USA)	Storage phosphor plate (manufacture not specified)	65 kVp, 7 mA, 0.125–0.32 s
Bayraktar (2022) [19]	Kirikkale University, Faculty of Dentistry, Turkey	Permanent teeth	Gendex Expert DC (Gendex Dental Systems, IL, USA)	–	65 kVp, 7 mA
Panyarak (2023) [32]	Chiang Mai University, Faculty of Dentistry, Thailand	Permanent teeth	Heliodent Plus (Dentsply Sirona, Charlotte, NC, USA)	Carestream 7600 Phosphor Plate (Carestream, Rochester, NY, USA) and VistaScan Imaging Plates Plus (DÜRR DENTAL, Bietigheim-Bissingen, Germany)	–
Estai (2022) [24]	Dental clinics in Perth, Australia	Age: 18–60	Australia	Photostimulable phosphor plate (not specified)	Australia
Lee (2021) [28]	Yonsei University, Conservative Dentistry, Korea	Permanent teeth	Kodak RVG 6200 digital radiography system with CS 2200 (Carestream, Rochester, NY, USA)	–	–
Baydar (2023) [17]	Eski¸sehir Osmangazi University, Faculty of Dentistry, Turkey	–	ProX periapical X-ray unit (Planmeca, Helsinki, Finland)	ProScanner Phosphor Plates and Scanning System (Planmeca, Helsinki, Finland)	60 kVp, 2 mA, 0.05 s
van Nistelrooij (2024) [33]	Dutch dental practice-based research network, The Netherlands	15–88 years	–	–	–
		202 females, 179 males, and 2 unknown
Ayan (2024) [13]	Kirikkale University, Faculty of Dentistry, Turkey	Not considered	Gendex Expert DC (Gendex Dental Systems, IL, USA)	–	65 kVp, 7 mA
Bayati (2025) [16]	Tehran University of Medical Science, Faculty of Dentistry, Iran	Permanent teeth	Owandy dental X-ray machine	Photostimulable phosphor plates, and a Sordex Digora scanner	60 kVp, 6 mA, 0.32 s

Authors	AI models	Images, n, training/total	Annotators, n/ experience	Lesion prevalence	Lesion type	Diagnostic performance of AI models
b Included studies on model architecture, dataset characteristics, annotation details, lesion prevalence in the training model, and diagnostic performance metrics of the AI models
Ayhan (2024) [14]	YOLOv7-AP-CBAM	1,000/1,170	2/>4 years	0.24	Interproximal	F1: 0.82, recall: 0.83, precision: 0. 87
Karakuş (2024) [26]	YOLOv8	−/860	2/>3 years	–	Overall caries	F1: 0.95, recall: 0.93, precision: 0.98
					Interproximal
					D1	F1: 0.96
					D2	F1: 0.96
					D3	F1: 0.93
					Occlusal	F1: 0.95
					Secondary caries	F1: 0.96
de Frutos (2024) [23]	YOLOv5 RetinaNet EfficientDet (D0 and D1)	8,342/8,539	6/–	–	Interproximal	YOLOv5 performed best, except for the F1 for enamel caries (EfficientDet D1)
					Enamel	F1: 0.53, AP: 0.60 (at IoU 0.3), FNR: 0.15
					Dentine	F1: 0.59, AP: 0.62 (at IoU 0.3), FNR: 0.22
					Secondary lesions	F1: 0.56, AP: 0.68 (at IoU 0.3), FNR: 0.16
Azhari (2023) [15]	Ensemble U-Net (ResNet50, ResNext101, VGG19)	–/771	3/–	–	Interproximal adult
					Sound	IoU: 0.98
					Primary	IoU: 0.23
					Moderate	IoU: 0.19
					Advanced	IoU: 0.51
					Overall	Mean F1: 0.52, mean IoU: 0.54
					Interproximal Ped
					Sound	IoU: 0.97
					Primary	IoU: 0.08
					Moderate	IoU: 0.17
					Advanced	IoU: 0.25
					Overall	Mean F1: 0.44, mean IoU: 0.38
Ahmed (2023) [12]	U-Net ensemble (ResNet50, ResNext101, Inception-ResNet-v2)	443/554	2/–	–	Overall caries	Mean F1: 0.54, mean IoU: 0.55
					Modified-ICIDAS system
					Sound	IoU: 0.96
					Outer 1/2 enamel	IoU: 0.21
					Inner 1/2 enamel	IoU: 0.23
					Outer 1/2 dentine	IoU: 0.35
					Inner 1/2 dentine	IoU: 0.41
Panyarak (2022) [31]	YOLOv3	994/1,425	3/>7	0.42	ICCMS™ 2 class	At IoU 0.50
					Overall caries	Precision 0.77, recall: 0.82, F1: 0.79, AP:0.72
				0.34	ICCMS™ 4 class	At IoU 0.50
				0.035	Initial stage	Precision: 0.75, recall: 0.67, F1: 0.71, AP: 0.57
				0.046	Moderate stage	Precision: 0.62, recall: 0.53, F1: 0.57, AP: 0.36
					Extensive stage	Precision: 0.75, recall: 0.78, F1: 0.76, AP: 0.67
				0.08	ICCMS™ 7 class	At IoU 0.50
				0.12	Outer 1/2 enamel	Precision: 0.35, recall: 0.23, F1: 0.27, AP: 0.10
				0.13	Inner 1/2 enamel	Precision: 0.48, recall: 0.63, F1: 0.55, AP: 0.38
				0.035	Outer 1/3 dentine	Precision: 0.55, recall: 0.55, F1: 0.55, AP: 0.37
				0.026	Mid 1/3 dentine	Precision: 0.48, recall: 0.56, F1: 0.52, AP: 0.36
				0.02	Inner 1/3 dentine	Precision: 0.52, recall: 0.61, F1: 0.56, AP:0.39
					Pulp	Precision: 0.77, recall: 0.61, F1: 0.68, AP:0.52
Kunt (2023) [27]	YOLOv5 RetinaNet	2,992/3,989	1/5	–	Overall caries	Ensemble models
	Faster R-CNN EfficientDet
						F1: 0.79–0.80, Precision: 0.81–0.83
						Recall: 0.77–0.79, AP: 0.84–0.86 (at IoU 0.5)
						Individual models
						F1: 0.72–0.76, Precision: 0.71–0.77
						Recall: 0.71–0.79, AP: 0.74–0.83 (at IoU 0.5)
					Small lesion	Individual models at IoU 0.5
						AP: 0.66–0.79
					Medium lesion	AP: 0.78–0.85
					Large lesion	AP: 0.67–0.85
Vimalarani (2022) [34]	DG-LeNet	800/1,000	–/–	0.16	Overall caries	Accuracy: 0.99
						Sensitivity: 0.91, specificity: 0.99
					Premolars	Accuracy: 0.99
					Molars	Sensitivity: 0.93, specificity: 0.99
						Accuracy: 0.99
						Sensitivity: 0.92, specificity: 0.99
Mao (2021) [29]	AlexNet	195/278	3/>3	0.023	Overall caries	Accuracy: 0.90 (AlexNet)
	(GoogleNet					Accuracy: 0.87 (Google Net)
	Vgg19					Accuracy: 0.80 (Vgg19)
	ResNet50)					Accuracy: 0.83 (Resnet50)
Moran (2021) [30]	Inception	–/112	1/–	0.24	Interproximal	The inception model with a 0.001 learning rate performed best
	ResNet				Incipient	Precision: 0.72, recall: 0.87
						Specificity: 0.83, NPV: 0.93, AUC: 0.86
				0.13	Interproximal	Precision: 0.69, recall: 0.733
					Advanced	Specificity: 0.83, NPV: 0.86, AUC: 0.81
Bayrakdar (2022) [18]	VGG-16	–/621	2/>9	–	Overall caries	VGG-16 (detection)
						Sensitivity: 0.84, precision: 0.84, F1: 0.84
						U-Net (segmentation)
						Sensitivity: 0.81, precision: 0.86, F1: 0.84
					Overall caries	VGG-16 (detection)
					External dataset	Sensitivity: 0.77, precision: 0.78, F1: 0.78
						U-Net (segmentation)
						Sensitivity: 0.76, precision: 0.87, F1: 0.81
Forouzeshfar (2024) [25]	VGG16, VGG19, AlexNet, ResNet50	–/1,517	2/–	–	Overall caries	VGG19 performed best
						Accuracy: 0.91–0.94
						F1: 0.90–0.93, precision: 0.90–0.93
						Sensitivity: 0.91–0.95, specificity: 0.89–096
Cantu (2022) [20]	U-Net with EfficientNet-B5 encoder	3,293/3,686	4/>3 years	–	Overall caries	Accuracy: 0.80
						F1: 0.73
						Sensitivity: ≥0.70
Chaves (2024) [21]	Mask-RCNN with Swin Transformer backbone	340/425	2/2+3*	–	Interproximal	Interproximal
					Primary caries	Sensitivity: 0.74, precision: 0.68
						F1: 0.69; AUC: 0.81
					Secondary caries	Sensitivity: 0.70, precision: 0.75
						F1: 0.72; AUC: 0.80
Chen (2022) [22]	Faster R-CNN	818/978	3/>5 years	0.25	Interproximal	Accuracy: 0.87, F1: 0.74, precision: 0.77
					All caries	Sensitivity: 0.72, specificity: 0.93
					E1, E2	Sensitivity: 0.65, precision: 0.69
					D1	Sensitivity: 0.69, precision: 0.71
					D2 and D3	Sensitivity: 0.85, precision: 0.92
Bayraktar (2022) [19]	Modified YOLOv3	800/1,000	2/>10 years	0.16	Interproximal caries	Accuracy: 0.95, AUC: 0.87
						Sensitivity: 0.72, specificity: 0.98
						Precision: 0.87, NPV: 0.97
Panyarak (2023) [32]	ResNet-18 ResNet-50	2,333/2,758	3/>7	0.32–0.34	Overall caries	ResNet-50 performed best
	ResNet-101			0.035–0.036	ICCMS™ 4 class	Accuracy: 0.70–0.75, AUC: 0.88–0.91
	ResNet-152			0.046–0.047	Initial stage	Sensitivity: 0.78–0.83
					Moderate stage	Specificity: 0.54–0.67
					Extensive stage
				0.08	Overall caries	ResNet-152 performed best
				0.12	ICCMS™ 7 class	Accuracy: 0.53–0.62
				0.13	Outer 1/2 enamel	AUC: 0.84–0.87
				0.035	Inner 1/2 enamel	Sensitivity: 0.63–0.75
				0.026	Outer 1/3 dentine	Specificity: 0.30–0.34
				0.02	Middle 1/3 dentine
					Inner 1/3 dentine
					Pulp
Estai (2022) [24]	Faster R-CNN Inception-ResNet-v2	−/2,468	3/−	0.38	Interproximal	Accuracy: 0.87, F1: 0.87, recall: 0.89 precision: 0.86, specificity: 0.86
Lee (2021) [28]	U-Net	−/304	2/>5	–	Overall caries	Precision: 0.63, sensitivity: 0.65, F1: 0.64
Baydar (2023) [17]	U-Net	400/500	2/>3	–	Overall caries	Sensitivity: 0.82, precision: 0.95, F1: 0.88
Van Nistelrooij (2024) [33]	Mask-RCNN with Swin Transformer backbone	330/413	2/5	0.15	Secondary caries	AUC: 0.95, F1 score: 0.77, precision: 0.81
			4			Sensitivity: 0.74, specificity: 0.97
Ayan (2024) [13]	Modified YOLOv5	1,000/1,200	2/>10 years	0.26	Overall caries	mAP: 0.63–0.65
				0.06	All caries	mAP: 0.52–0.54
				0.20	Enamel caries	mAP: 0.74–0.77
					Dentine caries
Bayati (2025) [16]	YOLOv8	442/552	1/–	–	Interproximal caries	Precision: 0.84, recall: 0.80, F1: 0.82, FNR: 0.20
			1/>10 years		All caries	Precision: 0.96, recall: 0.96, F1: 0.96, FNR: 0.04
					Enamel caries	Precision: 0.80, recall: 0.73, F1: 0.77, FNR: 0.27
					Dentine caries

A dash (“–”) indicates missing information.

A star (“*”) indicates that cases with annotator discrepancies were reviewed by additional annotators.

F1, F1 score; AUC, area under the curve; FPR, false-negative rate; NPV, negative predictive value; AP, average precision; mAP, mean average precision; IoU, intersection over union.

Dataset Allocation

The number of total images eventually used across the studies varied widely, ranging from 112 in Moran et al. [30] to 8,539 images in de Frutos et al. [23], with a mean of 1,526 images and a median of 978. Not all initially collected images were used in the model training as some studies excluded images with poor quality during data preparation, while others did not.

There was a variation in the distribution of bitewing images used for training, validation, and testing across the included studies (Fig. 2). Six studies [13, 15, 24, 25, 28, 30] did not specify their data splits or provided unclear descriptions; therefore, they were excluded from this figure. Two of the 18 studies included provided data allocation based on the image number after augmentation [18, 26]. Most studies allocated a significant portion of their data to training, with smaller subsets reserved for validation and testing. On average, 81% of images were used for training (range 70–98%). Among the studies that specified a validation set [12, 16–18, 20, 21, 26, 27, 31–33], an average of 11% of the images was allocated for validation (range: 7–18%). For testing, the mean was 9% (range 2–14%).

Fig. 2.

Distribution of bitewing images used for training, validation, and testing in the studies that provided data allocation information. Two studies provided data split information for the augmented data, which is indicated with a “*.”

Data Augmentation

Most studies employed data augmentation techniques – including horizontal and vertical flipping, rotation, resizing, and cropping – to artificially increase the number of training images and enhance model generalization. Three studies did not specify whether they used augmentation [14, 25, 33]. Among the eight studies that reported image counts post-augmentation [16, 18, 19, 24, 26, 29, 30, 34], the datasets expanded significantly, ranging from 1,506 to 11,114, with an average increase of 9.1 times (range 2.7–30.9) compared to the original dataset. One study [28] did not use augmentation and instead utilized 12-bit depth images in the manufacturer’s raw bitewing format, potentially preserving more detailed image information.

Applied AI Models

Various deep learning models were employed, either individually or in combination, to automate detection, classification, and segmentation of caries. The majority were architectures primarily designed for classification tasks, such as ResNet50 and VGG19. For detection tasks, commonly used models included YOLO (you only look once) variants, such as YOLOv3 and YOLOv5, while U-Net variants were frequently used for segmentation tasks (Fig. 3).

Fig. 3.

Overview of the deep learning models used independently or in combination in the included studies, grouped by the primary automation tasks of detection, classification, and segmentation.

Training Time

The number of training epochs varied significantly among the studies, ranging from 100 epochs [29] to 5,000 epochs [19], with a mean of 488 epochs and a median of 200 epochs.

Annotator Characteristics

Subjective evaluation using bitewing radiographs was employed to establish the ground truth of caries presence and, in some studies, the depth of the lesion. One study provided an unclear description of annotation [34]. The remaining 22 studies established the ground truth for caries presence based on annotations from 1 to 6 annotators, with considerable differences in clinical experience (Table 2b). Most studies employed two or three annotators, typically experienced dentists, or oral radiologists with several years of clinical practice. The annotators’ role was not specified in one study [17] and unclear in another study, in which an automated segmentation via CanioCatch was applied [25].

Caries Prevalence in Image Material

Ten studies [13, 14, 19, 22, 24, 29–31, 33, 34] reported the prevalence of caries lesions in their datasets. The same image material was applied in two studies [31, 32]. Thus, only one study was included in calculating the mean and median of caries prevalence. Lesion prevalence ranged from 2.3% to 41.7%, with a mean prevalence of 24.9% and a median of 24.9%.

Diagnostic Performance

The last two columns in Table 2b summarize the lesion types analysed in each study, classified by affected tooth surface and lesion depth, along with the corresponding diagnostic performance of the deep learning models. While most studies evaluated overall caries diagnostic performance [12, 13, 17, 18, 20, 25, 27–29, 31, 32, 34], some provided a more detailed classification, reporting AI performance specifically for interproximal caries detection [14–16, 21–24, 26, 30], secondary caries detection [21, 23, 26, 33], and occlusal caries detection [26]. The performance of AI models was assessed using various performance metrics, as shown for each study in Table 2b. Some models outperformed or matched the performance of human dentists [18, 20, 22, 23, 27, 28, 30], particularly in detecting advanced lesions than initial lesions. Only one study presented the AI’s performance on an external dataset [18].

QUADAS-2 Bias Assessment

The QUADAS-2 evaluation indicated a generally high risk of bias across the included studies (Fig. 4 and Table 3). In the patient selection domain, all studies were rated high risk due to reliance on single-institution datasets and the exclusion of poor-quality or complex images, limiting generalizability. The index test domain frequently showed high or uncertain risk, often due to a lack of reported blinding to the reference standard. The reference standard domain consistently exhibited high risk as all studies relied solely on expert annotations without external gold standard validation. The flow and timing domain was often rated as uncertain due to insufficient reporting of data handling and timing procedures. Overall, these findings suggest that biases in patient selection, index test interpretation, reference standard usage, and study transparency may limit the confidence and applicability of the results.

Fig. 4.

QUADAS-2 summary plot.

Table 3.

QUADAS-2 risk of bias assessment

Authors	Domain	Risk of bias	Applicability concern	Narrative summary
Ayhan (2024) [14]	Patient selection	High	High	This study used 1,170 anonymized bitewing radiographs. However, radiographs with technical and positioning errors, or those with complex prosthetics, were excluded, which introduces a high risk of selection bias by not representing real-world variability and poses an applicability concern
	Index test	Unclear	Low	YOLOv7 was employed for tooth numbering and caries detection, achieving high accuracy. However, the study does not mention if blinding was used during test interpretation, raising concerns about observer bias
	Reference standard	High	Low	Annotations by two experienced professionals provided reliability, but the lack of an external gold standard, such as histological validation, raises the risk of reference standard bias
	Flow and timing	Unclear	Low	The timing between radiograph acquisition and labelling was not specified, and the handling of missing data is unclear, leading to uncertainty in this domain
Karakuş (2024) [26]	Patient selection	High	High	The study utilized 860 bitewing radiographs, excluding images with unknown grades or annotation issues. While the dataset is robust, exclusion criteria may introduce selection bias by omitting real-world cases, such as ambiguous or poor-quality images; this also poses an applicability concern
	Index test	Unclear	Low	The YOLOv8 model showed high precision (97.7%), recall (93.2%), and F1 score (95.4%). However, the lack of details regarding blinding during interpretation raises potential concerns about observer bias
	Reference standard	High	Low	Ground truth was established by six dental clinicians through consensus annotations. Despite this effort, the absence of an external gold standard, such as histological validation, introduces reference standard bias
	Flow and timing	Unclear	Low	The timing between radiograph acquisition and annotation is not detailed, nor is the handling of missing data, resulting in uncertainty in this domain
de Frutos (2024) [23]	Patient selection	High	Low	This study utilized 13,887 bitewing radiographs from the HUNT4 Oral Health Study, excluding images with unknown grades or annotation issues. Although it features a large dataset, the exclusion criteria could bias results by removing difficult or ambiguous cases, and this also poses an applicability concern
	Index test	Low	Low	Three state-of-the-art models (YOLOv5, RetinaNet, and EfficientDet) were evaluated, showing performance on par or better than clinicians. The index test aligns well with the study objective
	Reference standard	High	Low	Annotations by six dental clinicians and consensus annotation for the test set ensured reliability. However, no external gold standard (e.g., histological validation) was used, which could affect bias
	Flow and timing	Unclear	Low	The study lacks details on how data splits were performed at the patient level, risking potential leakage. The timing between data collection and annotation was also not described
Azhari (2023) [15]	Patient selection	High	High	This study used 771 bitewing radiographs, split into adult and paediatric datasets. Exclusion criteria, such as poor image quality or identifiable patient information, likely introduced selection bias by excluding cases representative of real-world variability; this also poses an applicability concern
	Index test	Unclear	Low	Semantic segmentation models trained on U-Net architecture with transfer learning showed high accuracy for advanced caries but struggled with primary and moderate caries. The lack of clarity about blinding during the interpretation raises concerns about observer bias
	Reference standard	High	Low	Ground truth labelling by two calibrated examiners was reliable, with kappa >0.8. However, no external gold standard (e.g., histological validation) was used, introducing reference standard bias
	Flow and timing	Unclear	Low	No information was provided on the timing of radiograph acquisition and annotation. Handling of missing data was not explicitly discussed, raising uncertainty about this domain
Ahmed (2023) [12]	Patient selection	High	High	This study utilized 554 bitewing radiographs and excluded those with technical issues, overlapping images, or indirect restorations. These exclusions could introduce selection bias by limiting the generalizability of the findings to real-world clinical scenarios; this also poses an applicability concern
	Index test	Unclear	Low	The U-Net model, using three encoders (ResNet50, ResNext101, and Vgg19), achieved an F1 score of 0.535. However, the absence of information on blinding during test interpretation raises potential observer bias concerns
	Reference standard	High	Low	Two calibrated restorative dentists labelled the images according to a modified ICDAS protocol, but the lack of an external gold standard such as histological validation introduces potential reference standard bias
	Flow and timing	Unclear	Low	The timing between image acquisition and labelling was not reported, and no mention was made of how missing or excluded data were handled, resulting in an unclear risk of bias in this domain
Panyarak (2022) [31]	Patient selection	High	High	The study analysed 994 bitewing radiographs, but the exclusion of teeth with overlapping surfaces exceeding half of the enamel likely introduced selection bias as this does not reflect real-world variability in clinical practice; this also poses an applicability concern
	Index test	Low	Low	YOLOv3 showed acceptable performance for caries detection and classification under IoU50 and IoU75 thresholds. However, there was a slight performance reduction at the higher IoU threshold. The test design aligns well with clinical and research objectives
	Reference standard	High	Low	Caries annotations were conducted by three experienced radiologists using a consensus approach. However, the absence of an external gold standard, such as histological validation, increases the risk of reference standard bias
	Flow and timing	Unclear	Low	The study lacks detailed information on the timing of radiograph acquisition and annotation, as well as how missing or excluded data were managed, leading to uncertainty in this domain
Kunt (2023) [27]	Patient selection	High	High	This study analysed 3,989 bitewing radiographs from four X-ray units. The exclusion of low-quality images and inconsistent scaling of radiographs introduces potential selection bias as these exclusions do not fully reflect clinical variability; this also poses an applicability concern
	Index test	Low	Low	Multiple object detection CNNs (YOLOv5, RetinaNet, Faster R-CNN, EfficientDet) were evaluated, with the best-performing ensemble achieving an F1 score of 0.80 and AP of 0.86. The use of advanced models and ensembling aligns well with clinical objectives
	Reference standard	High	Low	Caries annotations were performed by a single expert using bounding boxes. The reliance on a single annotator without external validation (e.g., histological confirmation) increases the risk of reference standard bias
	Flow and timing	Unclear	Low	The study lacks detailed information on the timing of image acquisition and annotation, as well as handling of missing data, leading to uncertainty in this domain
Vimalarani (2022) [34]	Patient selection	High	High	This study analysed 1,000 bitewing radiographs, but the exclusion of poor-quality images and augmentation-based dataset creation introduce selection bias as it does not reflect the variability seen in real-world clinical data; this also poses an applicability concern
	Index test	Unclear	Low	The proposed DG-LeNet classifier showed high accuracy (98.74%) and sensitivity (91.37%) for detecting caries. However, the study lacks details about blinding during test interpretation, raising concerns about potential observer bias
	Reference standard	High	High	The ground truth was determined by faculty-provided annotations without external validation, such as histological confirmation. This introduces a high risk of reference standard bias
	Flow and timing	Unclear	Low	The timing between image acquisition, annotation, and testing was not explicitly described, and handling of missing data was not reported, leading to uncertainty in this domain
Mao (2021) [29]	Patient selection	High	High	This study analysed 278 bitewing radiographs and excluded low-quality images. Additionally, image augmentation was applied to balance the dataset, which may not reflect real-world clinical diversity, introducing selection bias; this also poses an applicability concern
	Index test	Low	Low	The modified AlexNet CNN achieved an accuracy of 90.3% for caries detection and 95.56% for restoration detection, demonstrating strong performance and alignment with the study’s clinical goals
	Reference standard	High	Low	Annotations were provided by three professional dentists, but no external gold standard (e.g., histological confirmation) was employed, which raises the potential for reference standard bias
	Flow and timing	Unclear	Low	The timing between image acquisition and annotation is not detailed, nor is the handling of missing data explicitly described, leading to uncertainty in this domain
Moran 2021 [30]	Patient selection	High	High	The study analysed 112 bitewing radiographs, with exclusion criteria such as dental implants, crowding, and malocclusion. These exclusions may not fully represent real-world variability, introducing a high risk of selection bias; this also poses an applicability concern
	Index test	Low	Low	The CNN models (ResNet and Inception) achieved promising accuracy for caries classification, with the best-performing model achieving 73.3% accuracy. The models align well with the study’s objectives and clinical requirements
	Reference standard	High	Low	Ground truth annotations were based on a single expert’s evaluation without external gold standard validation, such as histological confirmation, introducing reference standard bias
	Flow and timing	Unclear	Low	The study lacks detailed information on timing between radiograph acquisition and annotation, as well as how missing or excluded data were handled, leading to uncertainty in this domain
Bayrakdar (2022) [18]	Patient selection	High	High	The study used 621 anonymized bitewing radiographs, excluding radiographs with artifacts and resizing images for standardization. These exclusions may introduce selection bias by not fully representing real-world clinical scenarios; this also poses an applicability concern
	Index test	Low	Low	The VGG-16 model for caries detection and U-Net for segmentation achieved strong internal performance metrics (F1 score = 0.84) and promising external validation results (F1 score = 0.78), aligning well with clinical and research objectives
	Reference standard	High	Low	Ground truth was established through annotations by two experienced clinicians without external gold standard validation, such as histological data, introducing a high risk of reference standard bias
	Flow and timing	Unclear	Low	The timing of radiograph acquisition and annotation, as well as the management of excluded or missing data, was not detailed, resulting in uncertainty in this domain
Forouzeshfar (2024) [25]	Patient selection	High	High	The study analysed 713 bitewing radiographs, excluding low-quality images and performing data augmentation to increase the dataset to 6,032 images. These exclusions and augmentations may introduce selection bias as they do not fully reflect real-world variability; this also poses an applicability concern
	Index test	Low	Low	Four CNN architectures (VGG16, VGG19, AlexNet, ResNet50) were employed, with VGG19 achieving the highest accuracy (93.93%). The models align well with clinical requirements and the research objectives
	Reference standard	High	Low	Annotations were performed by a specialist dentist and a radiologist assistant, but no external gold standard validation (e.g., histological confirmation) was used, increasing the risk of reference standard bias
	Flow and timing	Unclear	Low	The timing of radiograph acquisition, annotation, and model evaluation is not clearly specified. Handling of missing data was not reported, resulting in uncertainty in this domain
Cantu (2022) [20]	Patient selection	High	High	The study analysed 3,686 bitewing radiographs but excluded those where assessment was deemed impossible (e.g., poor-quality images). This introduces selection bias as these exclusions may not reflect real-world clinical challenges; this also poses an applicability concern
	Index test	Low	Low	The U-Net CNN with an EfficientNet-B5 encoder achieved an accuracy of 0.80 and outperformed dentists in sensitivity (0.75 vs. 0.36). The model aligns well with the study’s objectives and demonstrates robustness for detecting both initial and advanced caries lesions
	Reference standard	High	Low	Annotations were conducted independently by three experienced dentists and later reviewed by a fourth. However, the absence of an external gold standard (e.g., histological validation) increases the risk of reference standard bias
	Flow and timing	Unclear	Low	The timing of radiograph acquisition and annotation was not explicitly reported, nor was the handling of missing data, leading to uncertainty in this domain
Chaves (2024) [21]	Patient selection	High	High	This study analysed 425 bitewing radiographs from a dental practice network. Exclusion of radiographs with inappropriate angulation or distortion introduces potential selection bias as these conditions are common in real-world practice; this also poses an applicability concern
	Index test	Low	Low	The Mask-RCNN architecture with Swin Transformer backbone demonstrated high performance for detecting primary and secondary caries lesions, with FROC AUCs of 0.806 (primary) and 0.804 (secondary). This aligns well with the study’s clinical objectives
	Reference standard	High	Low	Caries annotations were conducted in three rounds by graduate and PhD students, with final consensus by experts. However, the absence of an external gold standard, such as histological validation, increases the risk of reference standard bias
	Flow and timing	Unclear	Low	The study lacks details on the timing of radiograph acquisition and annotation or how missing data were managed, leading to uncertainty in this domain
Chen (2022) [22]	Patient selection	High	High	The study used 978 bitewing radiographs and excluded primary teeth, poor-quality images, and distorted radiographs. These exclusions, while improving data quality, introduce selection bias by limiting the real-world applicability; this also poses an applicability concern
	Index test	Low	Low	The faster R-CNN model achieved a sensitivity of 0.72 and specificity of 0.93, significantly outperforming postgraduate students. Its diagnostic accuracy and robustness across different lesion stages demonstrate strong clinical applicability
	Reference standard	High	Low	Annotations were conducted by two endodontic experts and one radiologist, but no external gold standard such as histological validation was used, leading to a high risk of reference standard bias
	Flow and timing	Unclear	Low	The study did not provide sufficient details on the timing of radiograph acquisition and annotation or the handling of missing data, resulting in uncertainty in this domain
Bayraktar (2022) [19]	Patient selection	High	High	The study analysed 1,000 bitewing radiographs, excluding images that lacked sufficient diagnostic quality. This exclusion may introduce selection bias by removing challenging cases seen in real-world clinical scenarios; this also poses an applicability concern
	Index test	Low	Low	The YOLOv3-based CNN demonstrated an accuracy of 94.59%, sensitivity of 72.26%, and specificity of 98.19%. The use of a state-of-the-art object detection model aligns well with the study’s objectives and shows strong applicability for clinical use
	Reference standard	High	Low	Caries lesions were annotated by two experienced dentists through consensus, but the lack of external validation with a histological gold standard increases the risk of reference standard bias
	Flow and timing	Unclear	Low	The study lacks detailed information on the timing of radiograph acquisition, annotation, and testing, as well as handling of excluded or missing data, leading to uncertainty in this domain
Panyarak (2023) [32]	Patient selection	High	High	The study analysed 2,758 annotated bitewing radiographs, excluding images with severe occlusal attrition or poor quality. These exclusions may limit real-world variability and introduce selection bias and poses an applicability concern
	Index test	Low	Low	Four ResNet models (18, 50, 101, 152) were employed for ICCMS™-RSS classification. ResNet-50 achieved the best balance of sensitivity (78%) and specificity (67.19%) in the 4-class system, aligning well with clinical applications
	Reference standard	High	Low	Ground truth was annotated by three radiologists using a consensus approach but lacked an external gold standard (e.g., histological validation), increasing the risk of reference standard bias
	Flow and timing	Unclear	Low	The study lacks detailed information on timing of radiograph acquisition and annotation, and no mention of how missing data were managed, leading to uncertainty in this domain
Estai (2022) [24]	Patient selection	High	High	The study analysed 2,468 bitewing radiographs, excluding those with extensive restorations, primary teeth, or poor-quality images. These exclusions likely introduced selection bias by not representing real-world diversity in clinical scenarios; this also poses an applicability concern
	Index test	Low	Low	The 2-step DL system combining faster R-CNN for ROI detection and Inception-ResNet-v2 for classification achieved an F1 score of 0.87. The system demonstrated high diagnostic accuracy, aligning well with clinical and research objectives
	Reference standard	High	Low	Three experienced dentists annotated the images through a consensus approach. However, the lack of an external gold standard, such as histological validation, increases the risk of reference standard bias
	Flow and timing	Unclear	Low	The study does not provide details on the timing of image acquisition, annotation, or validation, nor does it specify how missing or excluded data were managed, leading to uncertainty in this domain
Lee (2021) [28]	Patient selection	High	High	The study used 304 training radiographs and 50 evaluation radiographs, excluding images with severe overlapping or poor quality. These exclusions, while improving image clarity, introduce selection bias and limit real-world applicability
	Index test	Low	Low	The U-Net CNN model achieved moderate performance with an F1 score of 64.14%. The model aligns well with clinical objectives and demonstrated utility in improving clinicians’ sensitivity, particularly for early caries detection
	Reference standard	High	Low	Annotations were performed by two experienced observers, with consensus reached for discrepancies. However, no external gold standard, such as histological validation, was used, increasing the risk of reference standard bias
	Flow and timing	Unclear	Low	The study lacks detailed information on the timing of radiograph acquisition and annotation. Handling of missing or excluded data was not explicitly described, leading to uncertainty in this domain
Baydar (2023) [17]	Patient selection	High	High	The study used 500 anonymized bitewing radiographs, excluding those with artefacts or poor quality. This exclusion likely introduced selection bias by not representing real-world clinical diversity; this also poses an applicability concern
	Index test	Low	Low	The U-Net model demonstrated strong performance in segmentation tasks with F1 scores ranging from 0.8818 (caries detection) to 0.9722 (root canal filling). These results indicate high diagnostic accuracy, aligning well with the study’s objectives
	Reference standard	High	Low	Ground truth annotations were performed by two clinicians with differing experience levels, but no external gold standard (e.g., histological confirmation) was used, increasing the risk of reference standard bias
	Flow and timing	Unclear	Low	The study does not provide details on the timing of radiograph acquisition and annotation or the handling of missing or excluded data, leading to uncertainty in this domain
van Nistelrooij (2024) [33]	Patient selection	High	Low	The study used 413 bitewing radiographs containing 2,612 restored teeth from 383 patients. Exclusion of radiographs with distortions or artifacts may introduce selection bias by omitting challenging real-world cases
	Index test	Low	Low	The Mask R-CNN with Swin Transformer backbone achieved high specificity (0.966) for detecting secondary caries, with lower sensitivity (0.737). The model’s ability to stage caries severity aligns well with clinical decision-making
	Reference standard	High	Low	Two annotators performed pixel-wise labelling, reviewed by senior dentists. The absence of an external gold standard, such as histological validation, increases the risk of reference standard bias
	Flow and timing	Unclear	Low	The study does not specify the timing of radiograph acquisition or annotation. Additionally, it does not provide details on the handling of missing or excluded data, leading to uncertainty in this domain
Ayan (2024) [13]	Patient selection	High	Low	The study used 1,200 bitewing radiographs, excluding canines, third molars, and images with restorations or secondary caries. Such exclusions may introduce selection bias by limiting the generalizability to typical clinical scenarios encountered in practice
	Index test	Low	Low	The modified YOLOv5 (YOLOv5xs) deep learning model demonstrated good detection performance (mAP: 65%, with 54% for enamel and 77% for dentin lesions). The model’s performance aligns well with clinical diagnostic objectives, indicating its strong applicability in educational contexts
	Reference standard	High	Low	Two specialist dentists labelled caries lesions through consensus. However, the absence of external validation (e.g., histological confirmation) introduces a high risk of reference standard bias
	Flow and timing	Unclear	Low	Details regarding the timing of radiograph acquisition and annotation procedures were insufficiently described. Additionally, there was no explicit information provided about how excluded or missing data were handled, leading to uncertainty within this domain
Bayati (2025) [16]	Patient selection	High	Low	The study used 552 bitewing radiographs, excluding images with poor quality, severe distortion, or significant overlapping proximal surfaces. Such stringent exclusions introduce selection bias by omitting more challenging real-world cases that clinicians typically encounter
	Index test	Low	Low	The YOLOv8 deep learning model achieved strong performance metrics, notably high precision (96.03%) and recall (96.03%) for enamel caries, demonstrating the model’s suitability and strong clinical applicability for caries detection tasks
	Reference standard	High	Low	Two annotators performed pixel-wise labelling, reviewed by senior dentists. The absence of an external gold standard, such as histological validation, increases the risk of reference standard bias
	Flow and timing	Unclear	Low	The study does not specify the timing of radiograph acquisition or annotation. Additionally, it does not provide details on the handling of missing or excluded data, leading to uncertainty in this domain

Risk of bias was assessed across four domains: patient selection, index test, reference standard, and flow and timing. A narrative summary is provided alongside the tabulated results.

mAP, mean average precision; AUC, area under the curve.

Discussion

The applied performance metrics vary significantly among studies, making it challenging to compare them. AI models for dental applications are primarily adapted from medical imaging diagnostics. In medical AI, sensitivity is often prioritized over specificity, particularly in conditions like cancer detection, where missing a diagnosis can have severe consequences. Many studies emphasize recall, precision, and F1 score, as medical datasets are often imbalanced, and decision thresholds can be adjusted to optimize performance.

However, dental AI requires a more balanced approach. In caries detection, both sensitivity and specificity are equally essential for ensuring accurate diagnosis and making informed treatment decisions. Overdiagnosis can lead to unnecessary interventions, while underdiagnosis may allow disease progression. Therefore, AI models for caries detection must carefully balance the detection of actual carious lesions while minimizing false positives. To ensure clinical relevance, studies should report a comprehensive set of diagnostic metrics, including recall (sensitivity), specificity, AUC-ROC, F1 score, and AP, to provide an accurate and meaningful evaluation of model performance in real-world dental practice.

The preference for architectures like ResNet and YOLO indicates their effectiveness in image recognition and object detection within dental radiography. These models excel at processing complex image patterns and detecting features indicative of caries, especially advanced lesions. However, studies have highlighted limitations of the AI models for detecting early-stage caries [12, 13, 15, 27, 31], while two studies reported lower performance in detecting dentinal caries [16, 30]. This discrepancy may partly stem from variations in the training data, such as the distribution of lesion stages used for model training. AI models tend to perform better when the dataset substantially represents the types of lesions being analysed. This underscores the need for larger, more diverse datasets that encompass caries at various stages. Models specifically tailored to recognize subtle early-stage demineralization can be challenging [3].

The included studies explored various deep learning architectures and training techniques for caries detection, segmentation, and severity classification, directly addressing practical needs. The development of these models relied on clinicians’ subjective evaluation for image annotation. The lack of a “gold standard” (histological validation) for caries detection may lead to lowered sensitivity for initial caries diagnostics [40] and potential bias. The number and expertise of annotators varied widely, with some studies not specifying qualifications, impacting the reliability of annotations and model training [41].

For deep learning models in caries diagnostics to be robust and generalizable, they must be trained on diverse, high-quality datasets that include images from various imaging systems, collected across multiple geographic locations, and representing a wide range of patient demographics and imaging conditions. Such diversity enhances the model’s ability to adapt to real-world clinical scenarios, improving reliability and diagnostic accuracy.

In this systematic review, several studies excluded radiographs of poor image quality [12, 14–17, 23, 24, 26, 28, 31–33]. Most studies relied on bitewing radiographs from a single institution or local clinics, resulting in a lack of demographic diversity in the training data (Table 2a). Only one study utilized radiographs collected from six institutes or research centers on a national scale [23]. However, no AI models were developed using internationally sourced, multicenter imaging datasets. Among the ten studies that reported the type of image receptor used for image acquisition, all employed storage phosphor plates, except for one study [27], which used three direct and one indirect acquisition systems (Table 2b). This raises concerns about the generalizability of the models to broader populations and varying clinical settings.

Data scarcity is common in medical imaging AI research [40]. Only a few studies provided data upon request; just two made the data fully open source [27, 33].

The reported diagnostic performance of deep learning models underscores their potential as adjuncts in the radiographic diagnosis of caries (Table 2b). A recent meta-analysis of 14 studies reported mean sensitivity and specificity values of 0.87 and 0.89 for caries detection using bitewing radiographs [6]. However, our systematic review highlights significant heterogeneity, including variations in radiographic acquisition, dataset characteristics, AI models and their tasks (detection, segmentation, and classification), annotation details, lesion types, and lesion prevalence. Additionally, limitations in the annotation process further restrict direct comparisons of diagnostic performance and impede universal generalizations. The lack of standardized methodologies and variability in reporting outcomes make it challenging to fully assess how these models compare to traditional diagnostic methods or each other. This inconsistency mirrors concerns raised in other systematic reviews of AI in medical imaging, where variability hampers evidence synthesis [42].

Risks of Bias

The risk of bias assessment revealed several concerns. Selection bias may result from using datasets from single institutions and excluding images with artefacts or poor quality, potentially producing models that are not generalizable to broader populations or environments [43]. Measurement bias arises from variability in annotation practices and the lack of standardization, introducing inconsistencies in the training data. The subjectivity inherent in caries diagnosis affects model performance. Algorithmic bias is another concern, where algorithms may inadvertently prioritize certain features, leading to missed subtle signs of caries or overemphasis on irrelevant patterns [44]. Although many of these biases are intrinsic to the field and do not fundamentally negate the findings, they should be carefully considered when interpreting the results.

Implications for Clinical Practice and Future Research

The high risk of bias and lack of external validation limit immediate applicability in diverse clinical settings. Models trained on homogeneous datasets may not perform adequately across different populations or imaging conditions. Therefore, these AI tools should currently be viewed as complementary aids rather than replacements for professional judgment. Ethical considerations, such as patient privacy, data security, and transparency in AI decision-making processes, must be addressed before widespread adoption [45].

High-quality, adequately reported AI studies on caries diagnostics using bitewings are needed [6]. Future research should prioritize large, multi-institutional datasets that encompass diverse patient demographics, imaging conditions, and caries presentations [46]. Efforts should also focus on minimizing selection, measurement, and algorithmic bias through techniques such as cross-validation in the training process. Additionally, standardizing annotation protocols is also crucial; establishing clear guidelines for annotator expertise, the number of annotators, annotation procedures, and consensus-building methods will improve the quality and reliability of ground truth data [47].

Conclusion

Deep learning models show promise in detecting and classifying dental caries using bitewing radiographs. While the studies demonstrate high diagnostic performance, especially in detecting advanced lesions, significant challenges remain. The high risk of bias, lack of standardization, limited dataset diversity, and absence of external validation constrain generalizability and clinical applicability.

Acknowledgments

The article was written with the assistance of Claude, ChatGPT, and Google Gemini for advice on formulations and proofreading but not as a source of facts. Figures were generated using Microsoft Excel, Python (v3.12.1), and the Matplotlib library (v3.5.2).

Statement of Ethics

Ethical approval was not required for this systematic review; a Statement of Ethics is not applicable because this study is based exclusively on published literature.

Conflict of Interest Statement

The authors confirm that there are no known financial or personal relationships that could have potentially influenced the work reported in this paper.

Funding Sources

The University of Bergen provided funding for the project by allocating research resources.

Author Contributions

K.S.H.H. has contributed substantially to the concept and study design, data acquisition, analysis or interpretation of data, and manuscript writing and revision. N.R.G. has contributed substantially to the concept or design, analysis or interpretation of data, and manuscript revision. X.Q.S. has contributed substantially to the concept and study design, data acquisition, analysis or interpretation of data, and manuscript writing and revision.

Funding Statement

The University of Bergen provided funding for the project by allocating research resources.

Data Availability Statement

All data analysed in this study are included in this published article and its online supplementary information files. Further enquiries can be directed to the corresponding author. Thirteen included in this review explicitly commented on data availability [13, 16, 17, 22, 23, 25–28, 30, 31, 33, 34]. Of these, seven studies [16, 22, 23, 25, 26, 30, 34] offered access to their data upon request. Two studies [27, 33] made their data fully available as open source.