Deep learning for tooth identification and numbering on dental radiography: a systematic review and meta-analysis

Abstract

Objectives

Improved tools based on deep learning can be used to accurately number and identify teeth. This study aims to review the use of deep learning in tooth numbering and identification.

Methods

An electronic search was performed through October 2023 on PubMed, Scopus, Cochrane, Google Scholar, IEEE, arXiv, and medRxiv. Studies that used deep learning models with segmentation, object detection, or classification tasks for teeth identification and numbering of human dental radiographs were included. For risk of bias assessment, included studies were critically analysed using quality assessment of diagnostic accuracy studies (QUADAS-2). To generate plots for meta-analysis, MetaDiSc and STATA 17 (StataCorp LP, College Station, TX, USA) were used. Pooled outcome diagnostic odds ratios (DORs) were determined through calculation.

Results

The initial search yielded 1618 studies, of which 29 were eligible based on the inclusion criteria. Five studies were found to have low bias across all domains of the QUADAS-2 tool. Deep learning has been reported to have an accuracy range of 81.8%-99% in tooth identification and numbering and a precision range of 84.5%-99.94%. Furthermore, sensitivity was reported as 82.7%-98% and F1-scores ranged from 87% to 98%. Sensitivity was 75.5%-98% and specificity was 79.9%-99%. Only 6 studies found the deep learning model to be less than 90% accurate. The average DOR of the pooled data set was 1612, the sensitivity was 89%, the specificity was 99%, and the area under the curve was 96%.

Conclusion

Deep learning models successfully can detect, identify, and number teeth on dental radiographs. Deep learning-powered tooth numbering systems can enhance complex automated processes, such as accurately reporting which teeth have caries, thus aiding clinicians in making informed decisions during clinical practice.

Introduction

Deep learning is a type of artificial intelligence (AI) algorithm that mimics the cognitive functions associated with the human mind, performing complex tasks like visual pattern recognition, analytical problem-solving, and decision-making. It can automatically extract complex features from large datasets, transforming data into insights without extensive manual processing.^1,2 At its core, deep learning is powered by artificial neural networks (ANNs), which are computational models inspired by information processing in biological neural networks.³ ANNs can model complex non-linear relationships between inputs and outputs. Through exposure to labelled training data, the network learns to map inputs to outputs by tuning its internal parameters, enabling subsequent prediction on new unlabelled data.⁴ One of the most impactful neural network architectures for deep learning is the convolutional neural network (CNN). CNNs are specifically designed for processing high-dimensional data such as images through the use of convolution operations. They are composed of a series of convolutional and pooling layers which enable multi-level hierarchical feature learning.⁵ This allows CNNs to automatically learn spatial feature representations directly from raw input data. CNNs can effectively perform visual recognition tasks such as object detection (localizing and identifying objects in images), semantic segmentation (labelling object boundaries at the pixel level), and image classification (categorizing image contents into distinct classes).^6,7 These capabilities have enabled numerous medical imaging applications of deep learning, including tasks such as brain tumour segmentation,⁸ skin lesion classification,⁹ and diabetic retinopathy screening.¹⁰

In dentistry, deep learning has been applied to dental radiographs for uses like the detection of periodontitis,¹¹ caries,^12,13 and periapical radiolucent lesions.¹⁴ One essential application is accurate tooth numbering (the process of charting teeth and assigning them numerical labels), identification (the model’s ability to recognize that these numerical labels correspond to teeth), and segmentation (the process of categorizing each pixel or voxel in an image or volume into predefined classes) which provides localization and spatial referencing that is crucial for subsequent dental procedures, diagnosis, and forensics.^15–17 However, manually numbering and identifying teeth on dental radiographs is time-consuming and error-prone. Automated AI solutions based on deep learning for tooth numbering could significantly reduce clinician workload and stress while also minimizing errors that can occur from fatigue or other factors.^18,19

Recent studies have demonstrated the potential for deep learning techniques to enable accurate automated tooth numbering and identification using dental radiographs.^4,16,20–51 However, a comprehensive synthesis of this evidence has not yet been performed. The purpose of this study is to systematically explore articles and studies on tooth numbering and identification that have used deep learning approaches on dental radiographic images.

Methods

Protocol and registration

This systematic review was reported based on the preferred reporting items for systematic reviews and meta-analyses extension for the diagnostic test accuracy (PRISMA-DTA) guideline.⁵² The protocol was registered at PROSPERO [CRD42022354178]. The definitions, formulas, and details of all metrics and parameters used in the methodology of this article can be found in Table S1.

Eligibility criteria

This article addresses the following PICO question: What are the performances (outcome) of the deep learning approaches (intervention) for tooth numbering and identification in dental X-rays (population) in comparison with the reference standard (comparison)?

Studies were included based on the following inclusion criteria:

P: Dental radiographs of human subjects (panoramic, peri-apical, bitewing, and cone-beam computed tomography [CBCT]).

I: Tooth numbering and classification of teeth by deep learning models.

C: Human performance.

O: Reporting performance measurements (eg, accuracy, sensitivity, specificity, sensitivity, precision, F1-score, receiver operating characteristic [ROC] curve, and area under the curve [AUC]).

Study exclusion criteria were as follows: Studies without sufficient details on the datasets (sample size), studies without a clear deep learning model, studies not mentioning X-ray modality, studies that did not have separate variable considering for the model’s performance on tooth numbering and segmentation. Moreover, review articles and conference proceedings which did not have full text were excluded.

Information sources and search

Search was conducted in the following databases until October 28, 2023: Pubmed, Google Scholar, Cochrane library, Scopus, IEEE, arXiv, and medRxiv. Each database was searched with adapted keywords (Table 1). Moreover, a manual search was conducted on the reference list of included studies to find any missing articles.

Table 1.

Specific search query for each database.

Database	Keywords
Pubmed	(“artificial intelligence” [MeSH Terms] OR “computer vision” OR “deep learning” [MeSH Terms] OR “machine learning” [MeSH Terms] OR “artificial intelligence” OR “deep learning” OR “machine learning” OR “neural network” OR “diagnosis, computer assisted” [MeSH Terms]) AND (“radiography, dental, digital” [MeSH Terms] OR “radiography, panoramic” [MeSH Terms] OR “radiography, bitewing” [MeSH Terms] OR “photography, dental” [MeSH Terms] OR “Radiography” [MeSH Terms] OR Radiography OR “Cone-Beam Computed Tomography” [MeSH Terms]) AND (tooth OR dent*) AND (“classification” [MeSH Terms] OR classification OR number OR detect OR segment)
Google Scholar	allintitle: (“machine learning” OR “artificial intelligence” OR “deep learning”) AND (“dental” OR “tooth”) AND (“segmentation” OR “numbering” OR “charting”) AND (“dental radiography” OR “Panoramic” OR “bitewing” OR “periapical”)
Scopus	TITLE-ABS-KEY ( (“machine learning” OR “artificial intelligence” OR “deep learning”) AND (“dental” OR “tooth”) AND (“segmentation” OR “numbering” OR “charting”) AND (“dental radiography” OR “Panoramic” OR “bitewing” OR “periapical” OR “CBCT”) )
Cochrane	(“machine learning” OR “artificial intelligence” OR “deep learning” OR “computer vision” OR “neural network”):ti,ab,kw AND (dentistry OR tooth OR dental OR tooth numbering OR dental diagnosis OR dentistry):ti,ab,kw AND (“Radiography” OR “Panoramic” OR “Bitewing” OR “dental photography” OR “Diagnostic Imaging” OR “CBCT”):ti,ab,kw
ArXiv	AND all=“Machine Learning” OR “Artificial Intelligence” OR “Deep Learning”; AND all=dent* OR “tooth”; AND all=dental radiography OR dental photography OR Panoramic OR CBCT OR bitewing; AND all=segment* OR detect* OR number*
medRxiv	“Machine Learning” OR “Artificial Intelligence” OR “Deep Learning” AND “tooth” OR “dent”
IEEE	(“All Metadata”:“Machine Learning” OR “All Metadata”:“Artificial Intelligence” OR “All Metadata”:“Deep Learning” OR “All Metadata”:“computer vision” OR “All Metadata”:“neural network”) AND (“All Metadata”:“tooth” OR “All Metadata”:“dentistry” OR “All Metadata”:“teeth numbering”)AND (“All Metadata”:“Panoramic” OR “All Metadata”:“Bitewing” OR “All Metadata”:“dental photography” OR “All Metadata”:“CBCT”)

Study selection

We used Endnote X9 (Clarivate, Philadelphia, PA, USA) to manage data and merge references. After removing duplicates, 2 independent reviewers (F.T., M.L.) screened titles and abstracts, and the third reviewer (F.G.) resolved any discrepancies. Two investigators (F.T. and M.L.) assessed the full texts based on the inclusion and exclusion criteria and a third reviewer (F.G.) checked the disagreements.

Data collection and extraction

Two reviewers (P.R., Y.D.) collected data independently. In case of disagreements or discrepancies, it was resolved by a third reviewer (S.S.). The following items were extracted: Detailed bibliographic information (authors and publication date), data modality, data source (if the data are public or not), dataset size (training, validating, testing), inclusion and exclusion criteria (if any), labelling procedure, classification groups, deep learning task (segmentation/object detection/classification), pre-processing, augmentations, model structure, performance measurements, and outcomes.

Risk of bias and applicability

Two reviewers independently (F.T. and Y.D.) assessed the risk of bias using the QUADAS-2 tool.⁵³ QUADAS-2 list includes the following domains: patient selection, index test, reference standard, and flow and timing. Disagreements were resolved by a third reviewer (R.R.). Each domain was rated low, unclear, and high based on the information reported. If all domains were graded low, the study was considered to have low bias risk. If one or more domains were unclear or high, the study was rated high. “Patient selection” bias was high if limited information on the dataset was presented as well as unclear data split strategies and data leakage. A high risk of bias for the “Index test” domain was associated with poor reporting on test reproducibility, insufficient information about the model construction, and the lack of robustness analyses of the model. “Reference standard” bias was regarded as high if there was a lack of information on the reference standard definition or using only 1 examiner for establishing the reference test. “Flow and timing” bias was considered high if different reference standards were employed across the same study and inappropriate intervals between the index test and reference standard.¹²

Synthesis

We used diagnostic odds ratios (DORs) as pooled outcomes, calculated in the following manner:

$$\mathit{DOR}=\frac{\mathit{Sensitivity}\times \mathit{Specificity}}{\left(1 - \mathit{Sensitivity}\right)\times \left(1 - \mathit{Specificity}\right)}$$

A limited number of studies (n = 3) have reported the number of true positives, false positives, and false negatives, but more studies (n = 5) have reported sensitivity and specificity. Therefore, we only included studies that reported both specificity and sensitivity in the quantitative analysis. ROC curves were used to calculate AUC using summary ROC (sROCs). MetaDiSc and STATA 17 (StataCorp LP, College Station, TX, USA) were used to generate all plots. MIDAS and Metandi packages were used in STATA 17.0 for all analyses. Funnel plots analyze reviews for publication bias. This bias often leads smaller studies to show larger effect sizes compared with larger studies. To assess this bias, Deeks’ funnel plot is utilized. With increased trial size, studies are likely to converge around the true underlying effect size. This bias stems from a tendency to submit and publish studies with positive results, creating an asymmetric funnel plot when large-effect, positive result studies are sparse, especially among smaller studies.

Certainty of evidence

The Grading of Recommendations Assessment, Development and Evaluation (GRADE) methodology was employed to assess the certainty and dependability of evidence sourced from an extensive array of studies covering diverse tasks, within the scope of the meta-analysis.⁵⁴

Results

Study selection

The PRISMA flow chart (Figure 1) shows how the selection procedure was conducted in the present systematic review. Among 1618 searched studies, 68 studies were included for full-text screening. Forty articles were excluded based on the exclusion criteria and one was included in an additional manual search. Twenty-nine remaining articles were included for further steps.

Figure 1.

PRISMA flow chart for included studies.

Study characteristics

A summary of the included studies (n = 29) is given in Table 2. Included studies reported on different dental X-ray modalities (periapical, bitewing, panoramic, and CBCT). A total of 21 studies trained their deep learning model with panoramic radiographs,^{4,19,20,26,30,32–34,41,47–51,55–59} 3 with CBCT,^60–62 2 with periapical,^31,63 and 3 with bitewing.^18,64,65 Based on the deep learning task, object detection was the most used task (n = 19), followed by segmentation (n = 10) and classification only (n = 2). The most used model was Faster-RCNN (n = 11).^{4,18–20,26,27,31–33,51,55,56}

Table 2.

Summary of included studies.

Author, Year (Reference)	Data modality	Dataset size (train/validation/test)	Inclusion and exclusion criteria (if any)	Labelling procedure	Numbering classes (studies used different classes for numbering)	Deep learning task	Pre-processing	Augmentations	Model structure	Performance measurement	Outcome
Jang et al.60	CBCT (Panoramic reconstruction for 2D object detection and segmentation)	97 (66/-/31)	Exclusion: CT slices that didn’t contain teeth from the bottom	The teeth in the images was classified into 4 different categories, which are incisor, canine, premolar, and molar	4	Object detection Segmentation Classification	NA	NA	Custom-made one-stage object detection (object detection and classification) U-Net (segmentation) Loose and tight ROIs for 3D tooth segmentation	IOU (object detection)	60%
										AP (object detection)	88%
										Precision (identification)	98%
										Sensitivity (identification)	95%
										F1-score (identification)	96%
										Precision (2D segmentation)	100%
										Sensitivity (2D segmentation)	99%
										DSC (2D segmentation)	98%
										Precision (Loose and tight ROIs 3D segmentation)	98%
										Sensitivity (Loose and tight ROIs 3D segmentation)	96%
Miki et al.62	CBCT	52 (42/0/10)	NA	Classified in 7 tooth types: central incisors, lateral incisors, canines, first and second premolars, and first and second molars	7	Classification	Rotating, resizing, changing contrasts	Crop squash fill, half crop half fill, rotation, and intensity transformation	AlexNet	Accuracy	89%
Li et al.61	CBCT	400 (200/0/200)	NA	Classified in 4 classes: incisor, canine, premolar, and molar	4	Classification	Converted to gray-level image, resampled to the size of 64 × 64	Image rotation gamma correctionnoise injectionrandom translationscalpingaffine transform	Seven-layer convolutional neural network	Sensitivity (incisor)	88%
										Sensitivity (canine)	86%
										Sensitivity (premolar)	84%
										Sensitivity (molar)	90%
										Sensitivity (average)	87%
Chung et al.34	Panoramic	818 (574/162/82)	NA	32 teeth and missing teeth were annotated using axis-aligned bounding boxes by experts	32	Object detection	CLAHE	NA	Custom-made multi-tasked CNN (ResNet-18, DLA-34, and HG-Stacked)	AP (ResNet-18)	81%
										AP (DLA-34)	80%
										AP (HG-Stacked)	77%
										mIoU (ResNet-18)	84%
										mIoU (DLA-34)	84%
										mIoU (HG-Stacked)	81%
										Precision (ResNet-18)	100%
										Precision (DLA-34)	100%
										Precision (HG-Stacked)	99%
										Sensitivity (ResNet-18)	97%
										Sensitivity (DLA-34)	97%
										Sensitivity (HG-Stacked)	96%
Mahdi et al.33	Panoramic	1000 (900/-/100)	Exclusion: severely broken teeth	Images were labelled manually by putting a rectangular bounding box around each tooth Data were manually labelled and divided into 8 classes	8	Object detection	NA	NA	Faster R-CNN (ResNet-50, ResNet-101)	Precision (ResNet-50)	97%
										Precision (ResNet-101)	99%
										Sensitivity (ResNet-50)	98%
										Sensitivity (ResNet-101)	98%
										F1-score (ResNet-50)	98%
										F1-score (ResNet-101)	98%
										mAP (ResNet-50)	97%
										mAP (ResNet-101)	98%
Oktay et al.	Panoramic	100 (80%/-/20%) 10-fold cross-validation	NA	Classified in 3 classes: incisors, premolars, and molars	3	Classification	Determining the mouth gap, mouth quarters rotated and mirrored, affline transportation of the potential tooth areas	NA	Modified AlexNet	Accuracy (molar)	94%
										Accuracy (premolar)	92%
										Accuracy (canine and incisor)	92%
Prados-Privado et al.30	Panoramic	2230 (80%/-/20%)	Inclusion: adults older than 18 years Exclusion: edentulous patients, images with temporary teeth and poor definition, images with removable prostheses, or images with only the presence of implants, computerized axial tomography, and radiography with overlap or objects out of the imaging plane.	Each image was revised by 2 examiners with more than 3 years of experience in general dentistry.	32	Object detection	NA	NA	Matterport mask RCNN (ResNet101)	Accuracy	93.83%
Muramatsu et al.41	Panoramic	100 (75/-/25)	NA	Data were classified by a dental radiologist in 4 classes: incisor, canine, premolar, and molar	4	Object detection Classification	NA	Rotation, horizontal flip	CNN DetectNet based on GoogLeNet (employed in tooth detection) ResNet-50 (employed in classification)	Sensitivity (object detection)	96%
										Accuracy by single sized model for tooth type (classification)	87%
										Accuracy by multi sized model for tooth type (classification)	93%
										Accuracy by single sized model for tooth condition (classification)	97%
										Accuracy by multi sized model for tooth condition (classification)	98%
Mahdi et al.55	Panoramic	1000 (900/100) 10-fold cross-validation	Inclusion: optimization algorithm Exclusion: severely broken teeth	Images were labelled by putting a rectangular bounding box around each tooth with proper roots and shape	32	Object detection Classification	NA	Flip, random crop	Faster R-CNN, (ResNet-50, ResNet-101)	mAP (ResNet-50)	96%
										mAP (ResNet-101)	96%
Motoki et al.26	Panoramic	1000(800/-/200)	Inclusion: All teeth with roots	Images were manually labelled using rectangular boundary box around per tooth (32 categories) including third molar	32	Object detection Classification	NA	NA	Faster R-CNN	Sensitivity (Faster R-CNN)	98%
										Sensitivity (optimization)	97%
										Precision ((Faster R-CNN)	92%
										Precision (optimization)	96 %
										Accuracy (Faster R-CNN)	92%
										Accuracy (optimization)	94%
										F1 score (Faster R-CNN)	95%
										F1 score (optimization)	97%
Estai et al.32	Panoramic	591 (90/10/-)	Inclusion: age over 18 Exclude: primary teeth, poor quality images, mixed dentition, heavy restorations , ridge, implants, or many metal crowns	Ground truth (SIANNO) by 3 qualified dentists with over 10 years of clinical experience, creating a bounding box for each tooth (32 categories)and assigned a tooth number as per the FDI tooth numbering system	32	Object detection classification	Automatic ROI detection using Unet	Height and width shift, rotation, and scaling	Faster R-CNN (object detection), VGG-16 (classification)	IoU (ROI detection)	71%
										Sensitivity (object detection)	99%
										Precision (object detection)	99%
										Sensitivity (classification)	98 %
										Specificity (classification)	100%
										Accuracy (classification)	100%
										Precision classification)	98%
										F1 score (classification)	98%
Kim et al.27	Panoramic	303 (253/0/50)	Inclusion: implant fixtures and crowns Exclusion: dislocated teeth, missing and late residual teeth	Labelled with dental objects (implant fixture, crown, tooth), and numbered by 3 dentists. Model can classify tooth as an incisor, canine, or molar	3	Object detection classification	Horizontal stretch, vertical stretch, and shear deformation	NA	Faster R-CNN (detection) and inception v3 (classification)	mAP (tooth detection) (IoU = 0.5)	97%
										mAP (implant fixture detection) (IoU = 0.5)	45%
										mAP (crown detection) (IoU = 0.5)	61%
										Sensitivity (classification)	75%
										Specificity (classification)	80 %
										Precision (classification)	84%
Tuzoff et al.19	Panoramic	1574 (1352/0/222)	Inclusion: Adults Exclusion: implant and fixed bridges	By 5 radiology experts, drawing bounding boxes around all teeth and provide a class label for each box with the tooth number (FDI system). data were classified in 32 classes.	32	Object detection classification	NA	Extending the region of cropped teeth	Faster R-CNN (detection), VGG-16 (classification)	Sensitivity (object detection)	99%
										Precision (object detection)	99%
										Sensitivity (classification)	98%
										Specificity (classification)	100%
Bilgir et al.20	Panoramic	2482 (1984/249/249)	Exclusion: radiographs with metal artifacts, position errors, movement, developmental anomalies, dental crowding, malocclusion, tooth rotation, anterior teeth with inclination, supernumerary and retained deciduous teeth and transposition of left canine and lateral	Two oral and maxillofacial radiologists with 10 years of experience, and 1 oral and maxillofacial radiologist with 3 years of experience. Data were classified in 32 classes using FDI system	32	Object detection	NA	NA	Faster R-CNN (inception v2)	Precision	97%
										F1-score	96%
										Sensitivity	96%
Kılıc et al.4	Panoramic	421 (329/46/46)	Inclusion: 5-7 years old pediatric patients Exclusion: Images with artifacts	Labelled by a pedodontist with 10 years of experience. Using FDI numbering system (20 classes, divided in 4 parts: each 8 classes)	20	Object detection	NA	NA	Faster R-CNN (Inception v2)	Precision	96%
										F1-score	97%
										Sensitivity	98%
Chandrashekar et al.56	Panoramic	1500(80/-/20)	NA	Dentists annotated data. Single model performed for 4 different tooth classes: molar, premolar, canine, and incisor	4 for single model 32 for collaboration model	Segmentation Object detection	NA	Flip, saturation, and contrast	Mask R-CNN, U-Net (segmentation) Faster R-CNN, YOLO-v5 (detection)	Accuracy (segmentation-collaborative)	99%
										F1-score (segmentation-collaborative)	99%
										mAP (segmentation-collaborative)	97%
										Accuracy (object detection-collaborative)	99%
										F1-score (object detection-collaborative)	99%
										mAP (object detection-collaborative)	98%
										Accuracy (segmentation-Mask R-CNN)	96%
										F1-score (segmentation- Mask R-CNN)	98%
										mAP (segmentation- Mask R-CNN)	95%
										Accuracy (object detection- Faster R-CNN)	91%
										F1-score (object detection- Faster R-CNN)	90%
										mAP (object detection- Faster R-CNN)	91%
										Accuracy (segmentation-U-Net)	97%
										F1-score (segmentation- U-Net)	94%
										mAP (segmentation- U-Net)	92%
										Accuracy (object detection-YOLO-v5)	99%
										F1-score (object detection- YOLO-v5)	100%
										mAP (object detection- YOLO-v5)	99%
Silva et al.38	Panoramic	543 (324/108/111)	Inclusion: 32 teeth, restorations and appliances Exclusion: Images contained implants and deciduous teeth	Annotated 543 images with number information	32	Segmentation Classification	Homogenized and optimized	Horizontal flip	Mask R-CNN, PANet, HTC, and ResNeSt	Accuracy (segmentation)(ResNeSt)	96%
										Specificity (segmentation)(ResNeSt)	98%
										Precision (segmentation)(ResNeSt)	93%
										Sensitivity (segmentation)(ResNeSt)	89%
										F1-score (segmentation)(ResNeSt)	91%
										mAP (segmentation) (ResNeSt)	69%
										Numbering mAP (ResNeSt)	72%
										Accuracy (segmentation) (HTC)	96%
										Specificity (segmentation) (HTC)	99%
										Precision (segmentation) (HTC)	94%
										Sensitivity (segmentation) (HTC)	86%
										F1-score (segmentation) (HTC)	90%
										mAP (segmentation) (HTC)	64%
										Numbering mAP (HTC)	7%
										Accuracy (segmentation) (PANet)	97%
										Segmentation Specificity (PANet)	99%
										Precision (segmentation) (PANet)	94%
										Sensitivity (segmentation) (PANet)	89%
										F1-score (segmentation) (PANet)	92%
										mAP (segmentation) (PANet)	71%
										Numbering mAP (PANet)	74%
										Accuracy (segmentation) (Mask R-CNN)	96%
										Specificity (segmentation) (Mask R-CNN)	99%
										Precision (segmentation) (Mask R-CNN)	94%
										Sensitivity (ssegmentation) (Mask R-CNN)	87%
										F1-score (segmentation) (Mask R-CNN)	90%
										mAP (segmentation) (Mask R-CNN)	97%
										Numbering mAP (Mask R-CNN)	70%
Oktay et al.58	Panoramic	478 (200/-/278)	Inclusion: restoration, appliances, missing teeth, and implant	The labels and bounding boxes of every tooth have been re-annotated by an expert. Teeth are grouped in 4 classes: incisors, canines, premolars, and molars.	4	Segmentation	NA	Horizontal flip, rotation, and adding noise	Mask R-CNN	F1-score (segmentation)	93%
										Precision (segmentation)	91%
										Sensitivity (segmentation)	95%
										IoU (segmentation)	82%
										Accuracy (detection)	98%
Yuksel, et al.54	Panoramic	600 (510/-/90)	Exclusion: Individuals younger than 12 years images with problems such as blurriness, superimpositions, distortions, and technician-related problems	Labelling was performed by intern dental students and then was validated by a professional endodontist	32	Object detection	Quadrants were detected by a separate model and each quadrant is cropped	Flip distortion	YOLO	AP (0.5:0.95)	47%
Zhang et al.63	Periapical	1000 (700/100/200)	NA	Data were annotated in 33 classes (32 teeth and the background)	32	Object detection	Graph-based method, contour detection method, level set method, and active contour without edges technique	NA	Cascade CNN	Sensitivity	96%
										Precision	96%
Tekin et al.64	Bitewing	1200 (100/-/200)	NA	Oral and Maxillofacial radiology specialists annotated data, FDI system was used	24	Segmentation	NA	NA	Mask R-CNN	Accuracy (numbering)	92%
										Precision (numbering)	94%
										Sensitivity (numbering)	95%
										F1-score (numbering)	93%
										Sensitivity (segmentation)	97%
										Precision (segmentation)	100%
Chen et al.31	Periapical	1250 (800/200/250)	NA	One expert dentist, rectangular bounding box was drawn. Teeth were classified in 32 classes	32	Object detection	NA	NA	Faster R-CNN DCNN	Precision (object detection)	99%
										Sensitivity (object detection)	98%
										Mean IOU	91%
										mAP	80%
										Numbering precision	92%
										Numbering sensitivity	91%
Majanga et al.65	Bitewing	11 148 (8361/-/2787)	NA	Randomly selected, tooth contour extraction algorithm was executed and manually categorized by 3 expert dentists. drawing bounded boxes of each tooth.	NA	Classification	Gaussian filter for image enhancement, noise removal, and canny edge detector for edge detection	Horizontal flip, vertical flip random crops, color jittering, shear, and random rotation	VGG-16 net	Accuracy	89%
Yasa et al.18	Bitewing	1125 (916/101/108)	Exclusion: Radiographs with metal imposition, cone-cut, position and motion artefacts, crown, bridges or implants	A radiology expert provided annotations according to FDI system	24	Object detection	NA	NA	Faster R-CNN (Google Net inception v2)	F1 score	95%
										Precision	93%
										Sensitivity	97%
Yilmaz et al.51	Panoramic	1200 (800/200/200)	Pediatric panoramic was excluded.	Three radiologists with 5, 10, and 20 years annotated the teeth	32 for numbering, and additional 4 classes for impacted teeth	Object detection	NA	NA	YOLO-V4, Faster R-CNN	Precision (YOLO-V4)	100%
										Recall (YOLO-V4)	99%
										F1-score (YOLO-V4)	100%
										Precision (Faster R-CNN)	94%
										Recall (Faster R-CNN)	91%
										F1-score (Faster R-CNN)	92%
Karaoglu et al.49	Panoramic	2702 (2231/-/471)	NA	Two dental radiologists labelled teeth	32	Segmentation	NA	NA	Mask R-CNN, Mask R-CNN + Heuristic algorithm	Precision (Mask R-CNN)	92%
										F1-score (Mask R-CNN)	92%
										Recall (Mask R-CNN)	91%
										Precision (Mask R-CNN + Heuristic algorithm)	96%
										F1-score (Mask R-CNN + Heuristic algorithm)	95%
										Recall (Mask R-CNN + Heuristic algorithm)	95%
Almalki et al.48	Panoramic	432(80%/-/20%)	NA	NA	32	Segmentation, object detection	NA	Noise addition and horizontal flipping	Cascade Mask R-CNN	Precision (object detection)	90%
										Precision (segmentation)	89%
Alam et al.47	Panoramic	1500 (1300/-/200)	NA	NA	32	Segmentation, classification	Resize	Cropping	VGG-16	Precision	89%
										F1-score	89%
										Sensitivity	88%
Xu et al.50	Panoramic	6046 (4232, 605, 1209)	Panoramic radiographs with strong motion blur, strong distortion, or wrong positioning were excluded	Two dental clinicians, each possessing over 5 years of experience annotated the teeth. An additional dentist, with more than 15 years of experience thoroughly reviewed all the annotations	32	Segmentation, classification	Resize	NA	Hybrid task cascade-based	Precision	97%
										Recall	97%

Included studies are categorized by the data modality and the publication year.

Abbreviations: AP = average precision; ANN = artificial neural network; AUC = area under curve; BPANN = back propagation artificial neural network; CNN = convolutional neural network; CNN-EL = combination of CNN and ensemble learning (EL); DCNN = deep convolutional neural network; DCRF = dense conditional random field, DLA = deep learning accelerator; DLC = deep learning contouring; DSC = dice similarity coefficient, FCM = fuzzy C-means; FDI = Federation Dentaire Internationale; KNN = K-nearest neighbors; mIoU = mean intersection over union; MLP = multilayer perceptron; NA = not available; NP = not public; PNN = probabilistic neural network; PSO = particle swarm optimization; ROC = receiver operating characteristic curve; ROI = region of interest.

Regarding the numbering classes, 16 studies considered all 32 teeth in the model,^{19,20,26,32,34,47–51,56,57,59,63} followed by 4 studies with 4 classes (incisor, canine, premolar, and molar),^41,58,60,61 2 studies with 24 classes (excluding third molar),^18,64 2 studies with 8 classes (central incisor, lateral incisor, canine, first premolar, second premolar, first molar, second molar, and third molar),^30,33 2 studies with 3 classes (1 study classified them in incisor, premolar, and molar and the other to incisor, canine, and molar),²⁷ 1 study with 20 classes (primary dentition),⁴ and 1 study with 7 classes (excluded third molars).⁶²

Risk of bias and applicability

QUADAS-2 assessment tools were used to evaluate the included studies (Figure 2). The included studies reported low bias in flow and timing due to the use of dental X-rays in deep learning models. Regarding the index text domain, 17 studies were classified as low risk of bias. The patient selection was mostly categorized as low bias (n = 23). The reference standard was classified as low risk of bias for 15 studies.

Figure 2.

Risk of bias assessment of included studies based on each domain.

Table 3.

Risk of bias and applicability assessment for each domain considered for each included study.

Green, yellow, and red bars represent studies with low, unclear and high risk of biases, respectively.

Results of individual studies

According to all included studies, deep learning had an accuracy range of 81.8%-99% in tooth numbering and segmentation and a precision range of 84.5%-99.94%. Additionally, sensitivity ranged from 75.5% to 98%, specificity ranged from 79.9% to 99, sensitivity was reported from 82.7% to 98%, and the F1-score ranged from 87% to 98%. In only 6 studies, the deep learning model was less than 90% accurate.

Studies that used CBCT reported a range of 93.8%-98% for precision. Chen et al.²⁵ reported 93.8% for the FCN model with 25 CBCTs which contained primary teeth, supernumerary teeth, implants, and metal-restored teeth. Jang et al.⁶⁰ reported 98% using 97 CBCTs for deep learning-based hierarchical multi-step model.

Regarding panoramic radiography, 87.21%-94.32% was reported for accuracy in studies that used object detection. Chung et al.³⁴ reported 87.21% for a dataset of 818 panoramic radiographs using ResNet DLA and HG-Stacked model. Nonetheless, Oktay et al.⁵⁸ reported 94.32% for 100 panoramic radiographs with AlexNet architecture and CNN model. The accuracy range of 93.2%-99% was reported by studies using classification. Estai et al.³² reported 99% for a dataset of 591 (without primary teeth/implants/crowns/bridges) using U-Net, RCNN, and VGG-16 models. Nevertheless, Muramatsu et al.⁴¹ used the CNN DetectNet (based on GoogLeNet) and the ResNet-50 models on 100 images and reported 93.2%.

Studies utilizing periapical radiographs had a precision range of 91.4%-96.1%. Zhang et al.⁶³ observed 96.1% for a dataset of 1000 images with a Cascade CNN model. However, Chen et al. reported 91.4% for a Faster R-CNN model and 1250 periapical X-rays.

Data synthesis

We pooled data from 5 studies according to their sensitivity, specificity, and sample size.^{19,27,32,57,60} In Figure 3, the 95% CIs for sensitivity and specificity are plotted as a forest plot (Figure 4a). The average DOR of the pooled data was 1612 (95% CI 19-19 266), the sensitivity was 89% (77%-95%), the specificity was 99% (83%-100%), and the AUC was 96% (94%-98%). There was evidence of publication bias according to Deek’s funnel plot (Figure 4b) and statistical test (P = .08).

Figure 3.

Individual studies are represented by squares and lines, and 95% CIs are indicated by their 95% squares and lines. Pooled values are shown in diamonds. Statistical heterogeneity is represented by I² and Q by weighted sums of squares differences, respectively.

Figure 4.

(a) ROC curve summary. Each circle represents a different study, whereas the solid square represents the intersection point between sensitivity and specificity. (b) In order to assess publication bias, funnel plots are used.

Quality of evidence assessment

The studies incorporated into the meta-analyses exhibit a moderate level of certainty of evidence as indicated by GRADE (Table 4).

Table 4.

The outcomes derived from the process of Grading of Recommendations Assessment, Development, and Evaluation (GRADE) were classified into 4 distinct groupings, each encompassing the studies that were incorporated.

Outcome	No. of studies (No. of patients)	Study design	Factors that may decrease certainty of evidence						Certainty of evidence
			Risk of bias	Indirectness	Inconsistency	Imprecision	Publication bias	Other considerations
Detect, identify, and number teeth	5 studies 3107 samples	Diagnostic accuracy study	Seriousa	Not serious	Seriousb	Not serious	Strongly suspectedc	Very strong associationd	⊕⊕⊕◯ Moderate

^aTwenty-four studies have at least one high or unclear risk of bias.

^bInconsistency emerged due to the application of diverse deep learning tasks and classification classes.

^cPublication bias detected in Deek’s statistical test (P = .08).

^dVery large-effect size (DOR > 5).

Discussion

Accurate tooth identification and numbering is a fundamental first step in dentistry, establishing a spatial reference system to map diagnoses, treatments, and procedures to specific tooth locations. Various dental images are used for this purpose, including CT, CBCT, and conventional radiography like intraoral (periapical and bitewing) and extraoral (cephalometric and panoramic) images.⁶⁶ With the high patient volume in dental clinics, numerous images are taken daily for diagnostic purposes. Analysing these images is challenging, requiring extensive training to produce accurate analyses and reports. The process is time-consuming and prone to errors from fatigue, stress, and other factors.^33,57 Deep learning tools can aid dentists in decision-making, treatment planning, and saving time by automating the tooth numbering process.⁶⁷

This systematic review and meta-analysis demonstrate that deep learning models can successfully detect, identify, and number teeth on dental radiographs with high accuracy, achieving performance comparable to human dentists. The pooled analysis found that deep learning tooth numbering has a sensitivity of 89% and a specificity of 99%. By increasing analysis speed and accuracy, deep learning numbering offers several advantages for clinical practice. Large volumes of dental images can be processed quickly and efficiently to identify each tooth. This saves significant time otherwise spent on manual numbering, allowing clinical staff to dedicate more effort towards other patient care tasks. The structured data from automated numbering also enable the filing of digital dental charts, improving efficiency.⁴¹ Deep learning models provide consistent and standardized tooth numbering across different clinics, practices, and even countries. This improves communication effectiveness between dental professionals, since the spatial mapping of tooth locations is standardized.⁶⁷ It also aids accurate documentation sharing and referrals between providers.

However, there are several limitations to the existing literature that warrant consideration before clinical implementation. Most studies in this review had relatively small sample sizes, often using just a few hundred images, frequently sourced from a single clinic or geographic region. This risks overfitting deep learning models to the particular demographics of the training data, reducing generalizability to new patient populations. Xu et al. provide 1 example of a larger and more heterogeneous dataset, with 6046 panoramic radiographs collected from multiple institutions and dental clinics throughout China, incorporating various imaging devices and a range of dental anomalies.⁵⁰ However, the dataset was still limited to a single country. More diverse datasets from multiple institutions and countries are needed to improve accuracy across global populations with varying dental anatomy, disease patterns, and imaging equipment.

Most studies focused solely on permanent dentition in adults. Two examined primary or mixed dentition, limiting utility in paediatric practice.^4,50 Additional data on deciduous and transitional dentition would expand model applicability. Model performance was also rarely evaluated on implants, crowns, and bridges commonly found in patient images but potentially underrepresented in training data.⁵⁰ This could hamper detection rates. However, Kim et al. reported that implant fixtures were more accurately detected due to their similar shapes and sizes in the training data, while the varying shapes and sizes of crowns in the training data led to a lower detection rate.²⁷

Moreover, included studies used different tooth numbering systems and classes, complicating comparative assessments. Several excluded third molars due to variability in eruption and extraction. Others simplified to just 3-8 tooth classes rather than all 32 permanent teeth. Standardizing tooth numbering methodology would enable easier cross-study analyses. There were also potential sources of bias in some studies, most notably the reference standard domain. Many relied on just 1 examiner to establish the reference test, while best practice recommends multiple raters. More rigorous reference standards would reduce bias. Following standardized checklists like CLAIM for reporting AI studies and STARD-AI for diagnostic accuracy can further strengthen validity.^68,69

The deep learning models employed included CNNs, Faster R-CNNs, U-Nets, and ensemble methods. CNNs effectively extract hierarchical features from images through cascading convolutional layers. Faster R-CNN combines these convolutional feature extractors with region proposal and classification modules for localization and identification. U-Nets augment CNNs with skip connections between encoder and decoder steps, improving segmentation performance. Overall, 2-stage detection methods like Faster R-CNN were more commonly implemented than single-stage detectors across the studies reviewed. However, no single architecture emerged as clearly superior across all studies and tasks. Further research should continue examining diverse model architectures and libraries to determine the optimal approach.

The current study conducted a comprehensive and systematic review of existing studies on tooth numbering, allowing for comparison and synthesis, although it had some limitations. First, the scope of the study was limited to deep learning models and did not include all AI models. Second, only radiographs were considered and intraoral scans and photographs were not included. Additionally, it is possible that some non-English studies were not identified due to language restrictions during the search in the databases. As the data modalities varied among the included studies in the meta-analysis, it is essential to acknowledge the observed heterogeneity in the results. Notably, the majority of the included studies relied on panoramic radiographs, while only 1 study utilized CBCT (Panoramic reconstruction for 2D object detection and segmentation). This diversity in imaging modalities is likely a significant contributor to the observed heterogeneity in the findings. The studies included in the review utilized various metrics to evaluate the performance of the test set, such as accuracy, sensitivity, and specificity. However, a meta-analysis was not conducted for all the included studies due to heterogeneity and poor reporting quality. Most studies did not provide sufficient details, which made it difficult to conduct a detailed analysis for all the included studies and decreased the generalizability of the results.

Conclusion

Deep learning was successful in identifying and numbering teeth with high accuracy. In the future, deep learning systems may assist human observation of radiographs and improve performance by supporting clinicians in detecting and numbering teeth. Deep learning model needs to be explainable, interpretable, and generalizable to be integrated into daily clinical care.

Supplementary material

Supplementary material is available at Dentomaxillofacial Radiology online.

Funding

None.

Conflicts of interest

None.

Ethics approval

Not applicable.