Transformer-based models in dentistry: a systematic review Authors: Xuhan Zhu, Mingwei Zhou, Masahiro Shimogishi, Eriko Marukawa, Xin Zhu Published: 2026 ABSTRACT BackgroundTransformer-based architectures have rapidly gained prominence in medical imaging due to their ability to model long-range dependencies and global contextual information more effectively than convolutional neural networks. In dentistry, their applications have expanded across diagnostic, predictive, and generative tasks, yet no comprehensive synthesis has systematically evaluated their performance and clinical relevance. This systematic review provides an up-to-date assessment of Transformer-based models in dental imaging and diagnostics.MethodsA structured search of Medline, Embase, Web of Science, Scopus, and Cochrane databases was performed for studies published from 2020 to August 2025, following the PRISMA-DTA guidelines for diagnostic test accuracy systematic reviews and a PROSPERO-registered protocol (CRD420251142603). Eligible studies applied Transformer-based architectures to dental clinical tasks and reported quantitative performance outcomes. Two independent reviewers conducted article screening, data extraction, and bias assessment, resolving discrepancies through consensus.ResultsA total of 112 studies met the inclusion criteria. We found 91 hybrid convolutional neural network (CNN)–Transformer architectures and 23 pure Transformer models for dental image segmentation, classification, anomaly detection, and multimodal fusion tasks. In subgroup analyses, hybrid CNN–Transformer models consistently outperformed pure Transformer architectures. A significant 5–8% performance gap was observed between internal and external validation. These hybrid models demonstrated higher accuracy in tooth and anatomical segmentation, caries and lesion detection, orthopedic and orthodontic analysis, implant localization, and craniomaxillofacial assessment. Generative and predictive applications, such as 3D crown and bone reconstruction, dental age and sex estimation, artifact reduction, and implant position prediction, have achieved promising technical performance in controlled experimental settings, supporting further clinical translational research. Key advantages included enhanced global feature representation, improved robustness in heterogeneous imaging conditions, and the ability to incorporate multimodal inputs. However, major limitations remain: the scarcity of large annotated datasets, heterogeneous evaluation protocols, and limited prospective or real-world validation.ConclusionsTransformer-based models represent a significant methodological advance for dental AI, with superior technical performance in controlled experimental studies of dental imaging and diagnostic tasks. However, their clinical translation remains at an early stage. Broader clinical adoption requires not only standardized datasets, harmonized benchmarks and explainability frameworks, but also rigorous prospective multi-center clinical validation, regulatory and ethical evaluation, and real-world workflow integration assessment.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12911-026-03627-6. Background Oral and dental diseases remain a major global health burden, affecting an estimated 3.5 billion people worldwide, with disproportionate impact in low- and middle-income countries due to limited access to timely, affordable care [1]. Conditions such as dental caries [2], periodontal diseases [3], oral cancers [4], and congenital anomalies [5] like cleft lip and palate contribute to complications ranging from discomfort to life-threatening outcomes when left undiagnosed or untreated. Although technological advances have improved early detection and intervention, the increasing incidence of oral diseases highlights the need for more effective diagnostic solutions [6]. Dental image analysis plays a critical role in the diagnosis and management of oral diseases [7]. Technologies such as panoramic and periapical radiography [8, 9], cone-beam computed tomography (CBCT) [10], magnetic resonance imaging [11], and ultrasonography [12] have enhanced diagnostic capabilities. However, these modalities also present challenges, including noise, anatomical overlap, and variability in image quality, which hinder consistent interpretation [13]. Moreover, manual image analysis by clinicians is time-intensive and inherently subjective, leading to inter-observer variability and a risk of diagnostic errors. These limitations underscore the need for more standardized, efficient, and scalable diagnostic tools [14]. Artificial intelligence (AI), particularly deep learning, has emerged as a promising solution for automating and augmenting diagnostic workflows in medical and dental imaging [15]. Convolutional Neural Networks (CNNs) have led this progress since 2012 [16], demonstrating substantial success in tasks such as tooth classification [17], segmentation [18], lesion detection [19], endodontic and periodontal diagnosis [20], dental implant classification [21], and even forensic dental applications [22]. Models such as U-Net [23], introduced in 2015, have become foundational in medical image segmentation. However, CNNs are limited by their inductive bias towards local feature extraction, which restricts their ability to model long-range dependencies and global context, an essential requirement for complex image interpretation tasks [24]. Transformer-based models, whose core workflow is illustrated in Fig. 1, have emerged as a powerful alternative to CNNs. First developed for natural language processing (NLP), their self-attention mechanism computes pairwise relationships between all input elements simultaneously, enabling efficient modeling of long-range dependencies and global contextual information [25]. In 2020, Dosovitskiy et al. adapted this architecture to computer vision tasks, introducing the Vision Transformer (ViT) [26], which segments images into fixed-size patches, linearly embeds them, and processes them as sequences, analogous to words in a sentence. This design enables Transformers to model spatial and contextual relationships across the entire image without the locality constraints of convolutional filters.Fig. 1The typical workflow of the Transformer architecture proceeds through several sequential stages. Initially, input tokens are converted into dense vector representations via an embedding process and are subsequently augmented with positional encodings to preserve their sequential information, forming the initial input matrix. This matrix is then processed through the encoder stack, where it undergoes multi-head self-attention and feed-forward neural network operations. The resulting encoded representation, which encapsulates the contextual information of all input values, is passed to the decoder layers. Within the decoder, the data undergoes masked multi-head attention and encoder-decoder attention mechanisms before being processed by additional feed-forward networks. The final output from the decoder is transformed through a linear layer and a Softmax function to generate a probability distribution over the target vocabulary. Image by https://github.com/dvgodoy/dl-visuals/ CC BY The typical workflow of the Transformer architecture proceeds through several sequential stages. Initially, input tokens are converted into dense vector representations via an embedding process and are subsequently augmented with positional encodings to preserve their sequential information, forming the initial input matrix. This matrix is then processed through the encoder stack, where it undergoes multi-head self-attention and feed-forward neural network operations. The resulting encoded representation, which encapsulates the contextual information of all input values, is passed to the decoder layers. Within the decoder, the data undergoes masked multi-head attention and encoder-decoder attention mechanisms before being processed by additional feed-forward networks. The final output from the decoder is transformed through a linear layer and a Softmax function to generate a probability distribution over the target vocabulary. Image by https://github.com/dvgodoy/dl-visuals/ CC BY The flexibility of Transformers extends beyond single-modality image analysis. Their architecture can integrate diverse data types, such as medical images, electronic health records (EHRs), histopathological slides, genomic profiles, and physiological sensor streams, by jointly learning complex dependencies across modalities. This multimodal capability is particularly relevant for dentistry, where patient care often requires correlating radiographic findings with clinical records, histopathology, and procedural data. By leveraging self-attention to align and fuse information from multiple sources, Transformer-based models offer a pathway toward holistic, patient-specific diagnostic and prognostic systems. Furthermore, recent innovations have refined Transformer-based models for clinical image analysis tasks. Hybrid architectures that combine CNNs with Transformers exploit CNNs’ strength in extracting fine-grained local features while using Transformers to capture global structure and contextual dependencies. Architectures such as Swin Transformer [27], SegFormer [28], and TransUNet [29] incorporate hierarchical feature representation, window-based self-attention, and skip connections, making them more computationally efficient and better suited for high-resolution medical and dental images. In dentistry, these models have shown promise in tasks ranging from tooth segmentation in CBCT volumes to automated detection of caries, lesions, and implant positions [30], with several small-sample retrospective studies reporting performance comparable to that of dental clinicians in specific offline diagnostic tasks, such as classification [31], and anomaly detection [32]. Built on the core self-attention mechanism, Transformer architectures are inherently adaptable to both visual and textual data, with rapidly expanding applications across imaging and clinical NLP tasks in dentistry. This systematic review provides the first comprehensive synthesis of all Transformer-based models in dentistry, spanning ViT)/hybrid CNN-Transformer models for imaging, and Transformer-based large language models (LLMs) for clinical applications. We aim to map current evidence, characterize task-specific performance, and address barriers to clinical translation across these use cases. Recently, several review articles have explored the application of Transformer-based models in various domains of oral and dental health, including focused studies on stomatological imaging [33]. Moreover, some broader reviews of deep learning in dentistry have included discussions on Transformer architectures [6, 34–36]; however, no systematic review to date has summarized Transformer applications across modalities in dentistry. To fill this gap, we analyzed 112 eligible peer-reviewed articles, with findings structured into four core domains: image segmentation, detection/classification, generative/predictive imaging applications, and Transformer-based LLM clinical applications. This article follows a structured protocol: Sect. 2 elaborates on the methodology including the focused research question, eligibility criteria, database search strategy, data extraction and analysis, and stratified risk of bias assessment. Section 3 presents the research results covering literature search and selection, dental application classifications of Transformer models, quantitative performance synthesis, and systematic analysis of dataset and validation characteristics. Section 4 discusses the study implications, key limitations and targeted future research recommendations. Section 5 draws the conclusions, summarizing the application value of Transformer-based models in dental imaging and diagnostics and the core requirements for their clinical translation and real-world adoption. Methods Protocol development and focused question This systematic review was conducted in strict adherence to the PRISMA-DTA guidelines for diagnostic test accuracy systematic reviews [37]. The review protocol was registered and allocated with the CRD420251142603 in the PROSPERO International Prospective Register of Systematic Reviews hosted by the Centre of Reviews and Dissemination (University of York, National Institute for Health Research, United Kingdom). The focus question was “What is the trends, applications, strengths, and challenges associated with the use of Transformer-based models in dental applications?” Eligibility criteria Inclusion criteria Studies meeting all the following PIRD (Population-Index Test-Reference Standard-Outcome) criteria were included:Population (P): Clinical samples or patients across all fields of dentistry, covering all age groups, genders, and dental conditions (including healthy controls and patients with various oral diseases).Index Test (I): Models built on the core Transformer architecture, defined by self-attention as a core functional component. This includes:Vision Transformer architectures and hybrid CNN-Transformer models for dental imaging/3D data analysis;LLMs and NLP models for dental clinical applications.Reference Standard (R): Gold standard diagnoses or annotations established by at least one qualified dental clinician, including radiographic interpretations, histopathological diagnoses, clinical examination results, and expert consensus.Outcome (O): Quantitative performance metrics of the Transformer-based models in dental applications, including but not limited to accuracy (ACC), sensitivity AUC, specificity (SPE), Area Under the Curve (AUC), Intersection over Union (IoU), Positive/Negative Predictive Values (PPV/NPV), and Dice Similarity Coefficient (DSC).Study Design: All eligible observational studies (retrospective or prospective), diagnostic accuracy studies, clinical prediction model development and validation studies, and generative AI application studies. Population (P): Clinical samples or patients across all fields of dentistry, covering all age groups, genders, and dental conditions (including healthy controls and patients with various oral diseases). Population (P): Clinical samples or patients across all fields of dentistry, covering all age groups, genders, and dental conditions (including healthy controls and patients with various oral diseases). Index Test (I): Models built on the core Transformer architecture, defined by self-attention as a core functional component. This includes:Vision Transformer architectures and hybrid CNN-Transformer models for dental imaging/3D data analysis;LLMs and NLP models for dental clinical applications. Index Test (I): Models built on the core Transformer architecture, defined by self-attention as a core functional component. This includes:Vision Transformer architectures and hybrid CNN-Transformer models for dental imaging/3D data analysis;LLMs and NLP models for dental clinical applications. Vision Transformer architectures and hybrid CNN-Transformer models for dental imaging/3D data analysis; Vision Transformer architectures and hybrid CNN-Transformer models for dental imaging/3D data analysis; LLMs and NLP models for dental clinical applications. LLMs and NLP models for dental clinical applications. Reference Standard (R): Gold standard diagnoses or annotations established by at least one qualified dental clinician, including radiographic interpretations, histopathological diagnoses, clinical examination results, and expert consensus. Reference Standard (R): Gold standard diagnoses or annotations established by at least one qualified dental clinician, including radiographic interpretations, histopathological diagnoses, clinical examination results, and expert consensus. Outcome (O): Quantitative performance metrics of the Transformer-based models in dental applications, including but not limited to accuracy (ACC), sensitivity AUC, specificity (SPE), Area Under the Curve (AUC), Intersection over Union (IoU), Positive/Negative Predictive Values (PPV/NPV), and Dice Similarity Coefficient (DSC). Outcome (O): Quantitative performance metrics of the Transformer-based models in dental applications, including but not limited to accuracy (ACC), sensitivity AUC, specificity (SPE), Area Under the Curve (AUC), Intersection over Union (IoU), Positive/Negative Predictive Values (PPV/NPV), and Dice Similarity Coefficient (DSC). Study Design: All eligible observational studies (retrospective or prospective), diagnostic accuracy studies, clinical prediction model development and validation studies, and generative AI application studies. Study Design: All eligible observational studies (retrospective or prospective), diagnostic accuracy studies, clinical prediction model development and validation studies, and generative AI application studies. Exclusion criteria Studies with the following criteria were excluded:Studies published before 2020, non-English publications;Literature reviews, scoping reviews, unpublished manuscripts, conference proceedings, and book chapters without open access;Studies that did not report quantitative performance metrics of the Transformer-based models;Studies using non-Transformer core architectures (e.g., standalone CNNs), even with transformer-inspired modules;Studies that only compared non-Transformer models to Transformers without evaluating a Transformer-based intervention;Studies with incomplete or unverifiable data that could not be clarified through correspondence with authors. Studies published before 2020, non-English publications; Studies published before 2020, non-English publications; Literature reviews, scoping reviews, unpublished manuscripts, conference proceedings, and book chapters without open access; Literature reviews, scoping reviews, unpublished manuscripts, conference proceedings, and book chapters without open access; Studies that did not report quantitative performance metrics of the Transformer-based models; Studies that did not report quantitative performance metrics of the Transformer-based models; Studies using non-Transformer core architectures (e.g., standalone CNNs), even with transformer-inspired modules; Studies using non-Transformer core architectures (e.g., standalone CNNs), even with transformer-inspired modules; Studies that only compared non-Transformer models to Transformers without evaluating a Transformer-based intervention; Studies that only compared non-Transformer models to Transformers without evaluating a Transformer-based intervention; Studies with incomplete or unverifiable data that could not be clarified through correspondence with authors. Studies with incomplete or unverifiable data that could not be clarified through correspondence with authors. Information sources and search strategy We performed a systematic literature search across five databases: MEDLINE (via PubMed), Embase, Web of Science Core Collection, Scopus, and Cochrane Central Register of Controlled Trials (CENTRAL). The search yielded a total of 4862 articles across various sources: Scopus (1566), PubMed (1244), Web of Science (978), Embase (977), Cochrane (97). All searches were restricted to English-language publications only. The inclusion criteria, including (MeSH) terms, are specified in Table S2. A second electronic search was conducted in August 2025 to ensure up-to-date conclusions. Screening methods For managing the citations, Mendeley Version 2.135.0 was used. Nevertheless, the total number of citations was transferred to Rayyan’s online service (https://www.rayyan.ai) to ease the first selection process based on the article’s title and abstract. The screening process was performed by two independent reviewers (M.Z. and X.Z.). Full- text studies were considered if they followed the inclusion criteria or if they did not provide clear information regarding their methodology and purpose in the abstract. If disagreement was found at any stage of the selection process, it was solved through consensus with a third reviewer (X.Z.). The inter-reviewer reliability was registered by calculating the kappa correlation coefficient for both screenings. Data extraction and analysis Two independent reviewers (M.Z. and Xuhan Z.) performed duplicate, blinded data extraction for all included studies using a pre-standardized data extraction form. The following predefined data items were extracted from each eligible study: authors (year), the aim of the study, Transformer-based models used, the core features of models, dataset size and characteristics, primary performance metrics and corresponding values, and key outcomes or conclusions of the study. A third senior reviewer with clinical expertise in dentistry (Xin Z.) conducted a full review of the extracted data. All discrepancies and disagreements between the two independent reviewers during the data extraction process were resolved through consensus discussion with the third reviewer. For studies with incomplete or missing key data, we contacted the corresponding authors to request supplementary information and clarification. If the required information could not be obtained promptly, the study was excluded from the final data synthesis. For studies with overlapping patient cohorts or duplicate published results, only the most recent and comprehensive publication was included. For studies that evaluated multiple Transformer-based models in the same experiment, all relevant performance results were extracted and reported to ensure an objective and comprehensive synthesis. Risk of bias assessment We adopted a stratified, study-specific risk of bias (RoB) assessment strategy, tailored to the design and objectives of different types of included studies, using internationally validated, domain-specific tools to ensure the appropriateness and rigor of the evaluation. All RoB assessments were independently performed by two reviewers (M.Z. and Xuhan Z.). Discrepancies were resolved through consensus discussion with a third senior reviewer (Xin Z.), consistent with the study screening and data extraction workflow. Our stratified risk of bias assessment strategy is fully aligned with PRISMA-DTArecommendations for diagnostic test accuracy studies. For diagnostic accuracy studies focusing on dental anatomy segmentation, caries/lesion detection and classification, and pathological diagnosis (the majority of included studies), we used the QUADAS-AI tool [38], the AI-specific extension of the Cochrane-endorsed QUADAS-2 framework for diagnostic accuracy studies. This tool assessed RoB across four core domains: patient selection, index test, reference standard, and patient flow and timing. For each domain, we added AI-specific evaluation items, with a focus on data leakage risks, a prevalent and critical flaw in AI diagnostic research, including the rationality and timing of train-test data splitting, the presence of data leakage between training, validation, and test sets, control of overfitting, quality of reference standard annotation, and design of external validation. For clinical prediction model studies developing or validating clinical prediction models (e.g., dental age/sex estimation, implant position prediction, treatment outcome prediction), we applied the PROBAST (Prediction model Risk Of Bias ASsessment Tool) [39], the gold-standard framework for assessing bias in prediction model research. The assessment covered four key domains: participants, predictors, outcome, and analysis, with each domain rated for risk of bias and concerns regarding applicability. For LLM and generative model studies exploring LLM clinical applications and generative AI for dental clinical scenarios, we used a structured quality assessment checklist aligned with the CLAIM (Checklist for Artificial Intelligence in Medical Imaging) and TRIPOD-AI reporting standards [40, 41]. The checklist focused on model transparency, reproducibility of the analysis, clinical applicability, and completeness of reporting, with each item rated for risk of bias. The detailed results of the RoB assessment for each included study are provided in the Supplementary Materials (Table S3–S5). Results Search The selection process and flowchart are depicted in Fig. 2. The search strategy rendered a total of 4862 titles. After removing duplicate records, 3128 articles were included for screening. Subsequently, after passing title and abstract screening, 184 articles were sought for full-text analysis. From these, a final number of 112 articles were finally included for data extraction. Reasons for exclusion include not topic related and wrong objectives. Kappa Cohen’s Index for both title and abstract and full-text screenings were 0.927 and 0.899, respectively (P <0.001)Fig. 2PRISMA-DTA flow diagram for diagnostic test accuracy systematic reviews PRISMA-DTA flow diagram for diagnostic test accuracy systematic reviews As depicted in Fig. 3, there has been a marked and steady increase in the number of studies applying Transformer-based models within the field of dentistry over recent years. No such studies were reported in 2020; however, the research interest began to emerge in 2021 with a single publication, followed by a sharp rise in subsequent years—reaching 6 in 2022, 20 in 2023, and 41 in 2024. By early 2025, 44 additional studies had already been published, indicating sustained momentum in this area of research.Fig. 3Distribution of publications by years, categories of dental applications, anomaly/disease applications, and dental techniques Distribution of publications by years, categories of dental applications, anomaly/disease applications, and dental techniques Hybrid CNN–Transformer architectures were the mainstream of Transformer applications in the included studies, accounting for 81.3% (91 of 112) of eligible studies, while only 18.7% (21 of 112) used pure Transformer designs. This trend was consistent across all three major task categories. We further verified head-to-head performance comparisons with conventional CNNs under identical experimental settings, completed in 104 of 112 studies:For 26 segmentation studies: 25 reported Transformer-related models (24 using hybrid architectures) outperformed classic CNN baselines (e.g., U-Net) in core metrics (DSC, IoU).For 54 detection/classification studies: 51 completed controlled comparisons, with 48 reporting superior performance of Transformer-related models (44 hybrid designs) and 3 reporting comparable performance to state-of-the-art CNNs.For 32 generative/predictive studies: 27 completed CNN comparisons, all reporting better performance or unique clinical advantages of Transformer-related architectures (23 hybrid designs) over CNN baselines. For 26 segmentation studies: 25 reported Transformer-related models (24 using hybrid architectures) outperformed classic CNN baselines (e.g., U-Net) in core metrics (DSC, IoU). For 26 segmentation studies: 25 reported Transformer-related models (24 using hybrid architectures) outperformed classic CNN baselines (e.g., U-Net) in core metrics (DSC, IoU). For 54 detection/classification studies: 51 completed controlled comparisons, with 48 reporting superior performance of Transformer-related models (44 hybrid designs) and 3 reporting comparable performance to state-of-the-art CNNs. For 54 detection/classification studies: 51 completed controlled comparisons, with 48 reporting superior performance of Transformer-related models (44 hybrid designs) and 3 reporting comparable performance to state-of-the-art CNNs. For 32 generative/predictive studies: 27 completed CNN comparisons, all reporting better performance or unique clinical advantages of Transformer-related architectures (23 hybrid designs) over CNN baselines. For 32 generative/predictive studies: 27 completed CNN comparisons, all reporting better performance or unique clinical advantages of Transformer-related architectures (23 hybrid designs) over CNN baselines. Categories of dental applications used Transformer-based models Segmentation A total of 26 studies applied Transformer-based models for dental image segmentation (Table 1).Table 1Characteristics of the included studiesAuthors (year)Aim of the StudyTransformer model usedDataset featureDataset sizePrimary PerformanceMetrics and Values (%)Tooth SegmentationChen et al.(2023) [42]Tooth SegmentationCTA-UNetCBCT19,800 2D slicesDSC: 86.50IoU: 78.12Sheng et al.(2023) [43]Tooth segmentationSwin-UnetPanoramic radiographs100 panoramic radiographsF1-score: 63.72IoU: 46.89ACC: 88.52Wang et al.(2024) [44]Tooth segmentationTrans-VNetCBCTCBCT images from a total of 192 patients.DSC: 96.44IoU: 95.17Kanwal et al.(2023) [45]Tooth segmentationM-TransNetPanoramic radiographs1,500 panoramic radiographsACC: 97.25SPE: 97.65F1-Score: 93.47Li et al.(2024) [46]Tooth instance segmentationSPGTNetPanoramic radiographs56,34 panoramic radiographsmAP: 82.4ACC: 99.9Recall: 94.3F1-score: 94.8Hao et al.(2024) [47]Tooth segmentation and identificationSemiTNetPanoramic radiographsTSI15k: 16,317 imagesIoU: 94.41Dice: 95.45PRE: 94.74F1-Score: 95.90Ghafoor et al.(2023) [48]Tooth multiclass segmentationM-Net-like U-Net structurePanoramic radiographs540 annotated imagesACC: 97.26DSC: 91.02JI: 85.01PRE: 80.46Manzari et al.(2024) [49]Tooth segmentationDenUnetPanoramic radiographs2,043 annotated panoramic dental imagesACC: 95.4F1-score: 92.7Cao et al.(2024) [50]Tooth segmentationFlowgateUNetCBCT & microCT image24 extracted teeth, 11,969 slices in total+ 30 multi-organ CT image stacks (total 3779 slices)DSCCBCT: 90.69MicroCT: 83.86Synapse: 79.35Li et al.(2024) [51]Tooth segmentationS2AgScUNetPanoramic radiographsUFBA-UESC: 1500 total images across 10 categories; 425 images used in this studyIoU: 86.12Dice: 92.54PRE: 92.39F1-score: 92.54Tooth Segmentation (Cont.)Yin et al.(2023) [52]Tooth segmentationCoT-UNet++CBCTA self-constructed dataset of 20 groups with 300 CBCT images.Dice: 92.0695HD: 1.06 mmMa et al.(2025) [53]Tooth segmentationViT-B16Panoramic radiographs1,500 annotated dental panoramic X-ray images.IoU: 91.49Dice:94.54ACC: 97.42PRE: 94.59Wang et al.(2025) [54]Tooth segmentation across multimodal dental imagesTooth-ASAMCBCT & Panoramic radiographs & RGB (photos)4,983 CBCT scans + 2500 Panoramic radiographs + 409 videos of natural teethFor four datasets:Dice: 95.5/90.9/97.5/92.8IoU: 91.6/83.5/95.2/86.83D dental segmentationWu et al.(2024) [55]3D tooth segmentationPointRegion3D dental scans (CBCT)Labeled 916 dental models (403 maxillaries and 513 mandibles)OAcc: 97.109mIoU: 92.511Ahn et al.(2024) [56]3D dental image segmentationWCTNIntraoral 3D dental scans1,013 scansOA: 94.01mACC: 93.72mIoU: 88.71Zhuang et al.(2023) [57]Automatic teeth segmentation and labeling on 3D dental modelsAlignNetIntraoral 3D dental scans1,200 intraoral 3D scansDSC: 97.62Loc.: 99.13Lbl-T: 95.62Jin et al.(2023) [58]Automatic segmentation of 3D upper airway subregionsSwin-Unet3D dental scans (CBCT)50 cases, total 24,502 image slicesPRE: 94.25Recall: 98.44DSC: 96.29IoU: 92.8Alsheghri et al.(2024) [59]Dental arch segmentationArchSegIntraoral 3D dental arches1006 labeled upper and 984 lower 3D dental arches,DSC:Upper arches: 94.8Lower arches: 93.63D dental segmentation (con’t)Li et al.(2025) [60]3D segmentation and classification of jaw lesionsDense image-mask attention-guided Transformer network3D dental scans (CBCT)358 3D CBCT scans + 44-case CT datasetClassification ACC: 89.23Binary Segmentation Dice Score: 90.00Multi-Class Segmentation Dice Score: 79.06Krenmayr et al.(2025) [61]Investigates the effectiveness of four masked self-supervised learning (SSL) frameworks for 3D dental model segmentation, particularly tooth and brace segmentationPoint-BERT, Point-MAE, Point-GPT, Point-M2AEIntraoral 3D dental scans1. Pre-training dataset: over 4,000 unlabeled 3D dental models2. Tooth segmentation: 1,800 scans3. Braces segmentation: 52 scans used for binary segmentationTooth segmentation:Best Model: Pre-trained Point-MAEACC:93.26mIoU: 89.23Braces segmentation:Best Model: Pre-trained Point-MAEACC: 96.62mIoU: 93.58Dental Structures other than TeethChen et al.(2024) [62]Mandibular image segmentationCBTransPanoramic radiographsAnnotated 711 picturesACC: 90.39SEN: 87.84IoU: 91.87Lv et al.(2023) [63]Mandibular canal segmentationImproved Swin UNETRCBCT153 densely annotated primary datasets and 290 sparsely annotated auxiliary datasetsDice: 84.4IoU: 73.1clDice: 96.1HD95: 2.947 (mm)Liu et al.(2024) [64]Teeth, alveolar bone, maxillary sinus, and mandibular canal segmentationSwin-UNETRCBCT451 CBCT scans from 10 institutions and public datasetsDice: 95.1mIoU: 85.6HD: 1.90 mmASD: 0.27 mmLumetti et al.(2024) [65]Segmentation of the Inferior Alveolar Canal (IAC)MATAPCBCT443 annotated CBCT scans totalDice: 83.1IoU: 71.0Dental Structures other than Teeth (con’t)Gillot et al.(2022) [66]Segmentation of multiple anatomical skull structuresAMASSSCBCT volumetric scans of the head618 anonymized CBCT volumetric scansDice scores for Mandible/Skin/ Maxilla/Cranial base/Cervical vertebra: 96.2/97.1/85.3/78.8/76.0Tian et al.(2024) [67]Automatic segmentation of mandibular and maxillary cortical and cancellous bone structures3D UNesTCBCT155 dental CBCT scansDSC for mandibular cortical bone/mandibular cancellous bone/maxillary cortical bone/maxillary cancellous bone: 93.69/96.83/86.14/95.57ACC, Accuracy; ASD, average surface distance; AUC, Area Under the ROC Curve; CBCT, Cone Beam Computed Tomography; clDice, centerline Sørensen–Dice coefficient; Dice, Sørensen–Dice coefficient; DSC, Dice Similarity Coefficient; HD, Hausdorff distance; IoU, Intersection over Union; JI, Jaccard Index; Lbl-T, Labeling Accuracy; Loc, Localization Accuracy; Oacc, The overall accuracy; mACC, Mean Class Accuracy; mIoU, mean intersection over union; OPG, Orthopantomogram; PRE, Precision; SEN, Sensitivity; SPE, Specificit Table 1Characteristics of the included studiesAuthors (year)Aim of the StudyTransformer model usedDataset featureDataset sizePrimary PerformanceMetrics and Values (%)Tooth SegmentationChen et al.(2023) [42]Tooth SegmentationCTA-UNetCBCT19,800 2D slicesDSC: 86.50IoU: 78.12Sheng et al.(2023) [43]Tooth segmentationSwin-UnetPanoramic radiographs100 panoramic radiographsF1-score: 63.72IoU: 46.89ACC: 88.52Wang et al.(2024) [44]Tooth segmentationTrans-VNetCBCTCBCT images from a total of 192 patients.DSC: 96.44IoU: 95.17Kanwal et al.(2023) [45]Tooth segmentationM-TransNetPanoramic radiographs1,500 panoramic radiographsACC: 97.25SPE: 97.65F1-Score: 93.47Li et al.(2024) [46]Tooth instance segmentationSPGTNetPanoramic radiographs56,34 panoramic radiographsmAP: 82.4ACC: 99.9Recall: 94.3F1-score: 94.8Hao et al.(2024) [47]Tooth segmentation and identificationSemiTNetPanoramic radiographsTSI15k: 16,317 imagesIoU: 94.41Dice: 95.45PRE: 94.74F1-Score: 95.90Ghafoor et al.(2023) [48]Tooth multiclass segmentationM-Net-like U-Net structurePanoramic radiographs540 annotated imagesACC: 97.26DSC: 91.02JI: 85.01PRE: 80.46Manzari et al.(2024) [49]Tooth segmentationDenUnetPanoramic radiographs2,043 annotated panoramic dental imagesACC: 95.4F1-score: 92.7Cao et al.(2024) [50]Tooth segmentationFlowgateUNetCBCT & microCT image24 extracted teeth, 11,969 slices in total+ 30 multi-organ CT image stacks (total 3779 slices)DSCCBCT: 90.69MicroCT: 83.86Synapse: 79.35Li et al.(2024) [51]Tooth segmentationS2AgScUNetPanoramic radiographsUFBA-UESC: 1500 total images across 10 categories; 425 images used in this studyIoU: 86.12Dice: 92.54PRE: 92.39F1-score: 92.54Tooth Segmentation (Cont.)Yin et al.(2023) [52]Tooth segmentationCoT-UNet++CBCTA self-constructed dataset of 20 groups with 300 CBCT images.Dice: 92.0695HD: 1.06 mmMa et al.(2025) [53]Tooth segmentationViT-B16Panoramic radiographs1,500 annotated dental panoramic X-ray images.IoU: 91.49Dice:94.54ACC: 97.42PRE: 94.59Wang et al.(2025) [54]Tooth segmentation across multimodal dental imagesTooth-ASAMCBCT & Panoramic radiographs & RGB (photos)4,983 CBCT scans + 2500 Panoramic radiographs + 409 videos of natural teethFor four datasets:Dice: 95.5/90.9/97.5/92.8IoU: 91.6/83.5/95.2/86.83D dental segmentationWu et al.(2024) [55]3D tooth segmentationPointRegion3D dental scans (CBCT)Labeled 916 dental models (403 maxillaries and 513 mandibles)OAcc: 97.109mIoU: 92.511Ahn et al.(2024) [56]3D dental image segmentationWCTNIntraoral 3D dental scans1,013 scansOA: 94.01mACC: 93.72mIoU: 88.71Zhuang et al.(2023) [57]Automatic teeth segmentation and labeling on 3D dental modelsAlignNetIntraoral 3D dental scans1,200 intraoral 3D scansDSC: 97.62Loc.: 99.13Lbl-T: 95.62Jin et al.(2023) [58]Automatic segmentation of 3D upper airway subregionsSwin-Unet3D dental scans (CBCT)50 cases, total 24,502 image slicesPRE: 94.25Recall: 98.44DSC: 96.29IoU: 92.8Alsheghri et al.(2024) [59]Dental arch segmentationArchSegIntraoral 3D dental arches1006 labeled upper and 984 lower 3D dental arches,DSC:Upper arches: 94.8Lower arches: 93.63D dental segmentation (con’t)Li et al.(2025) [60]3D segmentation and classification of jaw lesionsDense image-mask attention-guided Transformer network3D dental scans (CBCT)358 3D CBCT scans + 44-case CT datasetClassification ACC: 89.23Binary Segmentation Dice Score: 90.00Multi-Class Segmentation Dice Score: 79.06Krenmayr et al.(2025) [61]Investigates the effectiveness of four masked self-supervised learning (SSL) frameworks for 3D dental model segmentation, particularly tooth and brace segmentationPoint-BERT, Point-MAE, Point-GPT, Point-M2AEIntraoral 3D dental scans1. Pre-training dataset: over 4,000 unlabeled 3D dental models2. Tooth segmentation: 1,800 scans3. Braces segmentation: 52 scans used for binary segmentationTooth segmentation:Best Model: Pre-trained Point-MAEACC:93.26mIoU: 89.23Braces segmentation:Best Model: Pre-trained Point-MAEACC: 96.62mIoU: 93.58Dental Structures other than TeethChen et al.(2024) [62]Mandibular image segmentationCBTransPanoramic radiographsAnnotated 711 picturesACC: 90.39SEN: 87.84IoU: 91.87Lv et al.(2023) [63]Mandibular canal segmentationImproved Swin UNETRCBCT153 densely annotated primary datasets and 290 sparsely annotated auxiliary datasetsDice: 84.4IoU: 73.1clDice: 96.1HD95: 2.947 (mm)Liu et al.(2024) [64]Teeth, alveolar bone, maxillary sinus, and mandibular canal segmentationSwin-UNETRCBCT451 CBCT scans from 10 institutions and public datasetsDice: 95.1mIoU: 85.6HD: 1.90 mmASD: 0.27 mmLumetti et al.(2024) [65]Segmentation of the Inferior Alveolar Canal (IAC)MATAPCBCT443 annotated CBCT scans totalDice: 83.1IoU: 71.0Dental Structures other than Teeth (con’t)Gillot et al.(2022) [66]Segmentation of multiple anatomical skull structuresAMASSSCBCT volumetric scans of the head618 anonymized CBCT volumetric scansDice scores for Mandible/Skin/ Maxilla/Cranial base/Cervical vertebra: 96.2/97.1/85.3/78.8/76.0Tian et al.(2024) [67]Automatic segmentation of mandibular and maxillary cortical and cancellous bone structures3D UNesTCBCT155 dental CBCT scansDSC for mandibular cortical bone/mandibular cancellous bone/maxillary cortical bone/maxillary cancellous bone: 93.69/96.83/86.14/95.57ACC, Accuracy; ASD, average surface distance; AUC, Area Under the ROC Curve; CBCT, Cone Beam Computed Tomography; clDice, centerline Sørensen–Dice coefficient; Dice, Sørensen–Dice coefficient; DSC, Dice Similarity Coefficient; HD, Hausdorff distance; IoU, Intersection over Union; JI, Jaccard Index; Lbl-T, Labeling Accuracy; Loc, Localization Accuracy; Oacc, The overall accuracy; mACC, Mean Class Accuracy; mIoU, mean intersection over union; OPG, Orthopantomogram; PRE, Precision; SEN, Sensitivity; SPE, Specificit Characteristics of the included studies Chen et al. (2023) [42] DSC: 86.50 IoU: 78.12 Sheng et al. (2023) [43] F1-score: 63.72 IoU: 46.89 ACC: 88.52 Wang et al. (2024) [44] DSC: 96.44 IoU: 95.17 Kanwal et al. (2023) [45] ACC: 97.25 SPE: 97.65 F1-Score: 93.47 Li et al. (2024) [46] mAP: 82.4 ACC: 99.9 Recall: 94.3 F1-score: 94.8 Hao et al. (2024) [47] IoU: 94.41 Dice: 95.45 PRE: 94.74 F1-Score: 95.90 Ghafoor et al. (2023) [48] ACC: 97.26 DSC: 91.02 JI: 85.01 PRE: 80.46 Manzari et al. (2024) [49] ACC: 95.4 F1-score: 92.7 Cao et al. (2024) [50] 24 extracted teeth, 11,969 slices in total + 30 multi-organ CT image stacks (total 3779 slices) DSC CBCT: 90.69 MicroCT: 83.86 Synapse: 79.35 Li et al. (2024) [51] IoU: 86.12 Dice: 92.54 PRE: 92.39 F1-score: 92.54 Yin et al. (2023) [52] Dice: 92.06 95HD: 1.06 mm Ma et al. (2025) [53] IoU: 91.49 Dice:94.54 ACC: 97.42 PRE: 94.59 Wang et al. (2025) [54] For four datasets: Dice: 95.5/90.9/97.5/92.8 IoU: 91.6/83.5/95.2/86.8 Wu et al. (2024) [55] OAcc: 97.109 mIoU: 92.511 Ahn et al. (2024) [56] OA: 94.01 mACC: 93.72 mIoU: 88.71 Zhuang et al. (2023) [57] DSC: 97.62 Loc.: 99.13 Lbl-T: 95.62 Jin et al. (2023) [58] PRE: 94.25 Recall: 98.44 DSC: 96.29 IoU: 92.8 Alsheghri et al. (2024) [59] DSC: Upper arches: 94.8 Lower arches: 93.6 Li et al. (2025) [60] Classification ACC: 89.23 Binary Segmentation Dice Score: 90.00 Multi-Class Segmentation Dice Score: 79.06 Krenmayr et al. (2025) [61] 1. Pre-training dataset: over 4,000 unlabeled 3D dental models 2. Tooth segmentation: 1,800 scans 3. Braces segmentation: 52 scans used for binary segmentation Tooth segmentation: Best Model: Pre-trained Point-MAE ACC:93.26 mIoU: 89.23 Braces segmentation: Best Model: Pre-trained Point-MAE ACC: 96.62 mIoU: 93.58 Chen et al. (2024) [62] ACC: 90.39 SEN: 87.84 IoU: 91.87 Lv et al. (2023) [63] Dice: 84.4 IoU: 73.1 clDice: 96.1 HD95: 2.947 (mm) Liu et al. (2024) [64] Dice: 95.1 mIoU: 85.6 HD: 1.90 mm ASD: 0.27 mm Lumetti et al. (2024) [65] Dice: 83.1 IoU: 71.0 Gillot et al. (2022) [66] Tian et al. (2024) [67] ACC, Accuracy; ASD, average surface distance; AUC, Area Under the ROC Curve; CBCT, Cone Beam Computed Tomography; clDice, centerline Sørensen–Dice coefficient; Dice, Sørensen–Dice coefficient; DSC, Dice Similarity Coefficient; HD, Hausdorff distance; IoU, Intersection over Union; JI, Jaccard Index; Lbl-T, Labeling Accuracy; Loc, Localization Accuracy; Oacc, The overall accuracy; mACC, Mean Class Accuracy; mIoU, mean intersection over union; OPG, Orthopantomogram; PRE, Precision; SEN, Sensitivity; SPE, Specificit Transformers have advanced medical image segmentation by capturing global context and long-range dependencies. In dentistry, they have been used across oral endoscope images, CBCT, periapical and panoramic radiographs, and 3D dental models, with CBCT and panoramic X-rays most frequently applied. Hybrid CNN–Transformer models dominate recent work. For CBCT, Chen et al. proposed CTA-UNet, combining convolutional encoders with Transformer layers, achieving a Dice Similarity Coefficient (DSC) of 86.52% on complex 2D slices [42]. Yin et al. introduced CoT-UNet++, integrating CoTNet and ViT, outperforming U-Net, UNet++, and TransUNet on a custom CBCT dataset [52]. Wang et al. developed Trans-VNet, which incorporated cross-attention into V-Net to address blurred boundaries and high inter-class similarity between teeth [44]. Together, these studies show CNN–Transformer hybrids substantially improve CBCT tooth segmentation. For panoramic X-rays, Sheng et al. used Swin-Unet with superior results over CNNs [43]. Kanwal et al. created M-TransNet by adapting SegFormer and Mask2Former, emphasizing Transformers’ strength in encoding global features [45]. Other teams, including Li, Ghafoor, Manzari, and Ma, also reported improved outcomes with hybrid designs [46, 48, 49, 53]. Hao et al. introduced SemiTNet, a semi-supervised framework reducing annotation requirements [47], while Li et al. pretrained SAM2 to build S2AgScUNet, a lightweight model effective on small datasets [51]. Beyond CBCT and panoramic images, Cao et al. applied FlowgateUNet to micro-CT and multi-organ CT data, showing strong generalizability [50]. Wang et al. advanced multimodal segmentation by integrating intraoral photographs, CBCT, periapical, and panoramic images, outperforming baselines across four benchmarks [54]. In 3D segmentation, Transformers excel due to their ability to model complex volumetric data. Wu et al. proposed a Region Partition Transformer for 3D point cloud tooth segmentation, achieving state-of-the-art results [55]. Ahn et al.’s WCTN, combining sparse convolution with Transformers, performed strongly on intraoral scans [56]. Zhuang et al. built a hybrid Transformer framework for fully automated 3D tooth segmentation and labelling [57], while Alsheghri et al. developed ArchSeg, based on Point Transformer V2, robust even on partial 3D arches [59]. Krenmayr et al. compared four masked self-supervised Transformer frameworks, finding Point-MAE most effective under limited labeled data [61]. Transformers have also improved segmentation of other dental structures. Chen et al. introduced CBTrans, a CNN–Swin hybrid, for mandibular segmentation on panoramic X-rays [62]. For CBCT, Lv et al. developed a model for mandibular canal segmentation with feature fusion and clDice loss [63]. Liu et al. combined Swin-UNETR for tooth instance segmentation with 3D-UNet for bone and nerve canals [64]. Lumetti et al. created MATAP, a memory-augmented Transformer incorporating spatial coordinates for mandibular canal segmentation [65]. Gillot et al. released an open-source UNETR-based tool for skull segmentation [66], and Tian et al. proposed a two-stage 3D UNesT system for cortical and cancellous bone segmentation [67]. Beyond dental anatomy, Transformers also support related regions: Jin et al. used Swin Transformer–U-Net for nasal and pharyngeal airway segmentation from CBCT [58]. Li et al. proposed an attention-guided Transformer for joint jaw lesion segmentation and classification, achieving strong results across both tasks [60]. Detection, classification, and localization A total of 54 studies applied Transformer-based models for diagnosing dental anomalies (Table 2).Table 2Characteristics of the included studies of the detection, classification and localizationAuthors (year)Aim of the StudyTransformer model usedDataset featureDataset sizePrimary PerformanceMetrics and Values (%)Tooth ClassificationGao et al.(2022) [31]Tooth classificationGrouped Bottleneck TransformerCBCT26 annotated CBCTACC: 91.3AUC: 99.7Sifat et al.(2024) [68]Classifying 3D tooth scans3DCNN + ViT3D dental models (CBCT)Total 34 de-identified tooth 3D modelsFor Dataset A/B:ACC: 90/100Wang et al.(2024) [69]Tooth point cloud registrationGeoSwin3D dental models (CBCT)37,828 pairs+ 11 sequences of LiDAR-scanned outdoor scenes.On Shining3D:RRE: 0.12°On Oral3:RRE: 0.197°On KITTI:RRE: 0.23°Zhao et al.(2025) [70]Accurate tooth detection and classificationRegion of Interest (RoI) transformerCBCTCBCT scans from 152 patientsACC: 98.48PRE: 97.29Recall: 97.21F1-score: 97.21mAP: 98.12Fernandes et al.(2025) [71]Compare the performance of three deep learning architectures in classifying various dental radiographic imagesViTs2D images derived from CBCT2,653 imagesACC:CNN: 0.71–0.99ViT: 0.74–0.98gMLP: 0.65–0.98AUC:CNN: 0.77–1.00ViT: 0.80–1.00gMLP: 0.73–1.00Liu et al.(2025) [72]Automatic tooth landmark detection on 3D dental modelsPoint Transformer V3 (PTv3)3D dental models (CBCT)3DteethLand: 340 annotated high-resolution 3D dental models with six categories of landmarks.mAP: 79.0mAR: 65.9Çelik et al.(2025) [73]Investigate whether deep learning–based super-resolution (SR) can enhance the classification accuracyESRTPanoramic radiographsTufts Dental Database: 6,400 cropped tooth imagesESRT + DenseNet (SF = 4):ACC: 92.3F1-score 92.2Anomalies/diseases detection, classification and localizationDental cariesFelsch et al.(2023) [30]Detection, classification and localization of caries and molar-incisor hypomineralization (MIH) classesSegFormer-B5Single-tooth photographs34,710 imagesF1-score: 97.7IoU: 95.9AP: 97.7ACC: 97.8Zhou et al. (2022) [32]Classification of the decayed teethT2S-TransformerPanoramic radiographs304 panoramic radiographsACC: 85.57PREC: 88.32Recall: 83.17F1-score: 85.67AUC: 92.23Jiang et al.(2021) [74]Detection and localization of dental caries and enamel hypomineralizationRDFNetDigital imageAnnotated 8,554 imagesPRE: 62.3Recall: 57.9mAP: 56.9FPS: 20.24Priya et al.(2024) [75]Segment and classify dental caries at an early stageATDUnet+ and ViT-MRDGRUPanoramic radiographs1,100 imagesACC: 96.59PRE: 96.55Ying et al.(2024) [76]Performance of four popular deep networks for dental caries detection compared with dentistsCompare with DETR,Trans-UNetPeriapical radiographs392 radiographs, after data augmentation, 1,730 imagesF1-scoreYOLOv5: 87; DETR: 86; UNet: 82; Trans-UNet: 80Youden Index:YOLOv5: 76; DETR: 73; UNet: 69; Trans-UNet: 64Ying et al.(2022) [77]Segmentation of dental cariesHybridPeriapical radiographs800 annotated images after augmentationDice: 74.87PRE: 74.43Anomalies/diseases detection, classification and localization (con’t)Dental caries (con’t)Neeraja et al.(2025) [78]Segment dental cariesDentC_BSFLNetPanoramic radiographs153 annotated OPGsACC: 98.53Dice: 95.93IoU: 92.19Chaves et al.(2024) [79]Detecting primary caries and secondary caries around restorationsMask R-CNN with Swin TransformerBitewing radiographs425 bitewingsFROC AUC: 80.6/80.4F1-score: 68.9/71.9for primary and secondary caries detection, respectivelyGao et al.(2024) [80]Early dental caries detectionGFDetEndoscopic imagesAnnotated 1318 annotated imagesPRE: 64.1Recall: 92.1F1-score: 75.6Huang et al.(2025) [81]Segmentation of dental cariesAEDD-NetPeriapical radiographs389 imagesDice: 76.8Jaccard: 64.4Various Dental DisordersAlsakar et al.(2024) [82]Classification of various dental disordersMobileNetV2 + Swin TransformerPanoramic radiographsAHE preprocessed 1272 dental radiographsPRE: 95.7SEN: 95.4SPE: 95.7ACC: 95.6Nour et al.(2025) [83]Diagnosis of dental diseasesYOLO8 and RT-DETR hybridPanoramic radiographs24,174 images at totalPRE: 90Recall: 79mAP50: 74ACC: 92Xu et al.(2025) [84]Automatic tooth instance segmentation and disease detectionUCL-NetPanoramic radiographs6,204 199 imaginesSegmentationmAP: 84.9Disease Detection:AP75: 64.5Li et al.(2025) [85]Segmentation of multiple oral pathologiesDual Swin TransformerPanoramic radiographsAnnotated 2,000 imagesDice scores for wisdom teeth, dental caries, and residual roots: 91.6/86.8/90.9Aljabar et al.(2024) [86]Automated diagnosis of dental conditionsViTPanoramic radiographAfter augmentation: 10,859 imagesACC: 96(ViT)94 (AlexNet + RBF SVM + PCA)Anomalies/diseases detection, classification and localization (con’t)Various Dental Disorders (con’t)Asif et al.(2025) [87]Diagnosis of oral and mouth diseases: Gingivostomatitis (GUM), canker sores (CaS), cold sores (CoS), oral lichen planus (OLP), mouth cancer (MC), oral thrush (OT) and oral cancer (OC)OralTransNetOral Disease ImageMOD dataset (5143 images, 7 classes)ACC: 99.03 (MOD), 97.83(Oral Cancer), 94.23 (Histopathology)F1 ≈ 99AUC: 99.98Peng et al.(2025) [88]Detection of oral diseasesOralformerOral Disease Image2602 image pairs: Dental plaque (1293), Calculus (606), Caries (703)mDice: 73.4 (Caries), 74.9 (Calculus), 74.8 (Plaques)Liu et al.(2025) [89]Improve dental disease classification (caries, deep caries, periapical lesions, and impacted teeth)ViT + Agent Attention mechanismPanoramic radiographsDENTEX2023 dataset: 1005 fully annotated panoramic dental X-ray imagesACC: 92.5PRE: 90.2Recall: 91.8F1-score: 91.0LesionsChen et al.(2024) [90]Tooth segmentation and lesion detectionSwin UNETRCBCT138 CBCT imagesDice: 85.12Ahmed et al.(2023) [91]Detecting lesions in noisy Optical Coherence Tomography (OCT) imagesTEDOCT imagesTotal 37,620 imaginesImproved the accuracy by 16–22% and IOU by 10%Panyarak et al.(2025) [92]Classifying common radiolucent jaw lesionsViT, CrossViT-15, CrossViT-18Panoramic radiographTotal annotated 208 imagines, expanded to 5149 training imagesBest performance (CrossViT-18 with ECAP):ACC: 72.6Anomalies/diseases detection, classification and localization (con’t)Lesions (con’t)Nistelrooij et al.(2025) [93]Evaluate the models for the automated detection and classification of osteolytic lesionsMask R-CNN, Mask R-CNN, Mask DINOPanoramic radiographs6,404 imaginesMask R-CNN + Swin-Tiny:F1-score (Well-Defined): 78.4AUC (Well-Defined): 88.1F1-score (Ill-Defined): 90.4AUC (Ill-Defined): 97.1Hadilou et al.(2025) [94]Classify histopathological images of oral epithelial dysplasia (OED)ViT-B16Histopathological images218 histopathological slide images, processed into total 7542 patchesACC: 94/97in 3-class/4-class scenarioThomas et al.(2025) [95]Precise localization and classification of dental cystsDentViT-RCNNPanoramic radiographs137 imagesACC: 99.99 (Segmentation)98 (Classification)CancerFlügge et al.(2023) [96]Detection of oral squamous cell carcinoma (OSCC)Swin-TransformerClinical intraoral photographs1406 imagesACC: 98.58F1-score: 98.57SEN: 98.57SPE: 98.59Vinayahalingam et al.(2024) [97]Detect and classify the oral potentially malignant disorders (OPMD) including leukoplakia and oral OLP and to distinguish them from OSCCMask R-CNN with a Swin TransformerIntraoral photographs4,161 images with pixel-level annotationsOSCC:F1-score: 85.2AUC: 97.4OLP:F1-score: 82.5AUC: 94.8.Leukoplakia:F1-score: 79.6 AUC: 93.8.Authors (year)Aim of the StudyTransformer model usedDataset featureDataset sizePrimary PerformanceMetrics and Values (%)Anomalies/diseases detection, classification and localization (con’t)Cancer (con’t)Maia et al.(2024) [98]Investigate the effectiveness of AI models for classifying oral cancer, including OSCC, dysplasia, and non-dysplasia.Compare with ViT, PiT, CoaTHistopathological imagesP-NDB-UFES: 3763 images, with 1930 Dysplasia, 1126 OSCC, 707 Non-dysplasiaBest:CNN: DenseNet-121 (BCC: 91.91, PRE: 91.93)Transformer: CoaT (BCC: 90.84)FSL: ProtoNet (BCC:88.42)Song et al.(2024) [99]Binary classification (suspicious vs. non-suspicious) of oral cancer images captured using mobile devicesSwin-B, ViT-BMobile-based oral cavity imagesTotal 2434 Annotated imaginesSwin Transformer:ACC: 88.7SEN: 90.5SPE: 87.0Impacted teethKüçük et al.(2025) [100]Detect impacted teethHybrid of YOLOv8 and RT-DETR,Panoramic radiographs407 panoramic radiographsmAP50: 98.3F1-score: 96PRE: 93.8Recall: 99.2FluorosisXu et al.(2025) [101]Diagnosis of dental fluorosisMLTrMRDental images200 dental imagesACC: 80.19F1-score: 75.79qw Kappa: 81.28Anomalies/diseases detection, classification and localization (con’t)Fluorosis (con’t)Gu et al.(2024) [102]Automatic grading diagnosis of dental fluorosis by segmentation and classificationLKU-Net, FusionDentNetTooth images131 patients’ imagesSegmentation:ACC: 95.25PRE: 94.06Classification:ACC: 93.3PRE: 85.7Dental plaqueSong et al.(2024) [103]Dental plaque segmentationCenterFormerOral endoscope images2,884 imagesACC: 86.51mIoU: 75.86mPA: 85.94Tez et al.(2025) [104]Detecting and segmenting dental plaque in the young permanent teethUNet TransformerIntraoral images506 intraoral images of anterior young permanent teeth from 31 children (ages 8–13)IoU: 78.45Dice score: 82.15OthersWang et al.(2025) [105]Detection of dens evaginatusResNet-18Panoramic radiographs1410 panoramic radiographsSEN: 86.0SPE: 92.0F1-score: 88.66AUC: 95.60Dujic et al.(2023) [106]Detecting periodontal bone loss (PBL)ViT-baseViT-largeBEiT-baseBEiT-largeDeiT-basePeriapical dental radiographs21,819 periapical radiographsACC: 85.2 (ViT-base)AUC: 0.918 (ViT-base)SEN: 91.3 (BEiT-large)SPE: 74.5 (ViT-base)PPV: 89.1 (ViT-base)NPV: 77.5 (BEiT-large)Çelebi et al.(2024) [107]Automatically detect the boundaries of pathological conditions and infections in the maxillary sinusRes-Swin-UNetCBCT298 imagesIoU: 84.71F1-score: 91.72PRE: 91.57Recall: 91.87ACC: 99.87Anomalies/diseases detection, classification and localization (con’t)Others (con’t)Tang et al.(2024) [108]Automatically classify midpalatal suture (MPS) maturation stagesCNN-enhanced ViTCBCT2,518 labeled CBCT imagesACC: 95.75PRE: 95.15Recall: 93.15F1-score: 94.13Vinayahalingam et al.(2024) [109]Detect carotid artery calcifications (CAC)Faster R-CNN w/ Swin-TPanoramic radiographsTotal annotated 6,404 panoramic radiographsPRE: 89.5Recall: 88.1F1-score: 88.8SPE: 98.7Takano et al.(2025) [110]Improve the automatic detection of calcification regionsTransFusePanoramic radiographs578 annotated dental panoramic radiographs, split into left/right halves, resulting in 1,156 imagesDice: 63.41IoU: 50.43Lu et al.(2023) [111]Craniomaxillofacial (CMF) anatomical landmark localizationCMF-NetCBCT150 subjects and each with 24 manually labeled skeletal landmarksMRE: 1.108 mmSDR: 79.28Resul et al.(2025) [112]Automated periodontitis diagnosisAPD-FFNetPanoramic radiographs337 panoramic dental radiographsACC: 94F1-score: 93.88AUC: 93.47Lin et al.(2025) [113]Detect whether the mandibular third molar compresses the inferior alveolar nerve (IAN)Transformer-based ConvNeXt_v2Panoramic radiographs3,295 Panoramic radiographs images (after augmentation).Segmentation ACC: 97.92Chang et al.(2025) [114]Automated orthodontic diagnosisSPMA networkCephalograms3634 lateral cephalogramsACC: 90.02MR: 71.38Anomalies/diseases detection, classification and localization (con’t)Others (con’t)Zhang et al.(2025) [115]Evaluate the performance of the ViT for classifying molars with or without furcation involvement (FI)ViTPanoramic radiographsExtracted 1,568 tooth images from 506 radiographs, 7,840 images after augmentationPRE: 98Recall: 92F1-score: 95ACC: 92AUC: 87Benakatti et al.(2025) [116]Evaluate DETR for identifying different dental implant systems from radiographic imagesDETRRadiographic images1138 radiographs (868 periapical radiographs, 270 orthopantomograms)PRE: 83Recall: 89F1-score: 82Çelik et al.(2025) [117]Explore the impact of various data augmentation techniques on automated segmentation of dental restorations (implants, prostheses, and fillings)SegFormer, Swin-T Knet, PCPVT-L UperNet, SVT-S FPNPanoramic radiographs204 panoramic radiographs from the Tufts Dental DatabaseIoU: 62–82Dice score: 75–90ACC, Accuracy; AP: Average Precision; ASD, average surface distance; AUC, Area Under the ROC Curve; CBCT, Cone Beam Computed Tomography; clDice, centerline Sørensen–Dice coefficient; Dice, Sørensen–Dice coefficient; DSC, Dice Similarity Coefficient; HD, Hausdorff distance; IoU, Intersection over Union; JI, Jaccard Index; mACC, Mean Class Accuracy; mAP: mean Average Precision; mAR: Mean Average Recall; MR: Match Ratio; MRE: Mean Radial Error; mIoU, mean intersection over union; NPV: negative predictive value; PA: Pixel Accuracy; PPV, Positive Predictive Value; PRE, Precision; SDR: Successful Detection Rate; SEN, Sensitivity; SPE, Specificity; Table 2Characteristics of the included studies of the detection, classification and localizationAuthors (year)Aim of the StudyTransformer model usedDataset featureDataset sizePrimary PerformanceMetrics and Values (%)Tooth ClassificationGao et al.(2022) [31]Tooth classificationGrouped Bottleneck TransformerCBCT26 annotated CBCTACC: 91.3AUC: 99.7Sifat et al.(2024) [68]Classifying 3D tooth scans3DCNN + ViT3D dental models (CBCT)Total 34 de-identified tooth 3D modelsFor Dataset A/B:ACC: 90/100Wang et al.(2024) [69]Tooth point cloud registrationGeoSwin3D dental models (CBCT)37,828 pairs+ 11 sequences of LiDAR-scanned outdoor scenes.On Shining3D:RRE: 0.12°On Oral3:RRE: 0.197°On KITTI:RRE: 0.23°Zhao et al.(2025) [70]Accurate tooth detection and classificationRegion of Interest (RoI) transformerCBCTCBCT scans from 152 patientsACC: 98.48PRE: 97.29Recall: 97.21F1-score: 97.21mAP: 98.12Fernandes et al.(2025) [71]Compare the performance of three deep learning architectures in classifying various dental radiographic imagesViTs2D images derived from CBCT2,653 imagesACC:CNN: 0.71–0.99ViT: 0.74–0.98gMLP: 0.65–0.98AUC:CNN: 0.77–1.00ViT: 0.80–1.00gMLP: 0.73–1.00Liu et al.(2025) [72]Automatic tooth landmark detection on 3D dental modelsPoint Transformer V3 (PTv3)3D dental models (CBCT)3DteethLand: 340 annotated high-resolution 3D dental models with six categories of landmarks.mAP: 79.0mAR: 65.9Çelik et al.(2025) [73]Investigate whether deep learning–based super-resolution (SR) can enhance the classification accuracyESRTPanoramic radiographsTufts Dental Database: 6,400 cropped tooth imagesESRT + DenseNet (SF = 4):ACC: 92.3F1-score 92.2Anomalies/diseases detection, classification and localizationDental cariesFelsch et al.(2023) [30]Detection, classification and localization of caries and molar-incisor hypomineralization (MIH) classesSegFormer-B5Single-tooth photographs34,710 imagesF1-score: 97.7IoU: 95.9AP: 97.7ACC: 97.8Zhou et al. (2022) [32]Classification of the decayed teethT2S-TransformerPanoramic radiographs304 panoramic radiographsACC: 85.57PREC: 88.32Recall: 83.17F1-score: 85.67AUC: 92.23Jiang et al.(2021) [74]Detection and localization of dental caries and enamel hypomineralizationRDFNetDigital imageAnnotated 8,554 imagesPRE: 62.3Recall: 57.9mAP: 56.9FPS: 20.24Priya et al.(2024) [75]Segment and classify dental caries at an early stageATDUnet+ and ViT-MRDGRUPanoramic radiographs1,100 imagesACC: 96.59PRE: 96.55Ying et al.(2024) [76]Performance of four popular deep networks for dental caries detection compared with dentistsCompare with DETR,Trans-UNetPeriapical radiographs392 radiographs, after data augmentation, 1,730 imagesF1-scoreYOLOv5: 87; DETR: 86; UNet: 82; Trans-UNet: 80Youden Index:YOLOv5: 76; DETR: 73; UNet: 69; Trans-UNet: 64Ying et al.(2022) [77]Segmentation of dental cariesHybridPeriapical radiographs800 annotated images after augmentationDice: 74.87PRE: 74.43Anomalies/diseases detection, classification and localization (con’t)Dental caries (con’t)Neeraja et al.(2025) [78]Segment dental cariesDentC_BSFLNetPanoramic radiographs153 annotated OPGsACC: 98.53Dice: 95.93IoU: 92.19Chaves et al.(2024) [79]Detecting primary caries and secondary caries around restorationsMask R-CNN with Swin TransformerBitewing radiographs425 bitewingsFROC AUC: 80.6/80.4F1-score: 68.9/71.9for primary and secondary caries detection, respectivelyGao et al.(2024) [80]Early dental caries detectionGFDetEndoscopic imagesAnnotated 1318 annotated imagesPRE: 64.1Recall: 92.1F1-score: 75.6Huang et al.(2025) [81]Segmentation of dental cariesAEDD-NetPeriapical radiographs389 imagesDice: 76.8Jaccard: 64.4Various Dental DisordersAlsakar et al.(2024) [82]Classification of various dental disordersMobileNetV2 + Swin TransformerPanoramic radiographsAHE preprocessed 1272 dental radiographsPRE: 95.7SEN: 95.4SPE: 95.7ACC: 95.6Nour et al.(2025) [83]Diagnosis of dental diseasesYOLO8 and RT-DETR hybridPanoramic radiographs24,174 images at totalPRE: 90Recall: 79mAP50: 74ACC: 92Xu et al.(2025) [84]Automatic tooth instance segmentation and disease detectionUCL-NetPanoramic radiographs6,204 199 imaginesSegmentationmAP: 84.9Disease Detection:AP75: 64.5Li et al.(2025) [85]Segmentation of multiple oral pathologiesDual Swin TransformerPanoramic radiographsAnnotated 2,000 imagesDice scores for wisdom teeth, dental caries, and residual roots: 91.6/86.8/90.9Aljabar et al.(2024) [86]Automated diagnosis of dental conditionsViTPanoramic radiographAfter augmentation: 10,859 imagesACC: 96(ViT)94 (AlexNet + RBF SVM + PCA)Anomalies/diseases detection, classification and localization (con’t)Various Dental Disorders (con’t)Asif et al.(2025) [87]Diagnosis of oral and mouth diseases: Gingivostomatitis (GUM), canker sores (CaS), cold sores (CoS), oral lichen planus (OLP), mouth cancer (MC), oral thrush (OT) and oral cancer (OC)OralTransNetOral Disease ImageMOD dataset (5143 images, 7 classes)ACC: 99.03 (MOD), 97.83(Oral Cancer), 94.23 (Histopathology)F1 ≈ 99AUC: 99.98Peng et al.(2025) [88]Detection of oral diseasesOralformerOral Disease Image2602 image pairs: Dental plaque (1293), Calculus (606), Caries (703)mDice: 73.4 (Caries), 74.9 (Calculus), 74.8 (Plaques)Liu et al.(2025) [89]Improve dental disease classification (caries, deep caries, periapical lesions, and impacted teeth)ViT + Agent Attention mechanismPanoramic radiographsDENTEX2023 dataset: 1005 fully annotated panoramic dental X-ray imagesACC: 92.5PRE: 90.2Recall: 91.8F1-score: 91.0LesionsChen et al.(2024) [90]Tooth segmentation and lesion detectionSwin UNETRCBCT138 CBCT imagesDice: 85.12Ahmed et al.(2023) [91]Detecting lesions in noisy Optical Coherence Tomography (OCT) imagesTEDOCT imagesTotal 37,620 imaginesImproved the accuracy by 16–22% and IOU by 10%Panyarak et al.(2025) [92]Classifying common radiolucent jaw lesionsViT, CrossViT-15, CrossViT-18Panoramic radiographTotal annotated 208 imagines, expanded to 5149 training imagesBest performance (CrossViT-18 with ECAP):ACC: 72.6Anomalies/diseases detection, classification and localization (con’t)Lesions (con’t)Nistelrooij et al.(2025) [93]Evaluate the models for the automated detection and classification of osteolytic lesionsMask R-CNN, Mask R-CNN, Mask DINOPanoramic radiographs6,404 imaginesMask R-CNN + Swin-Tiny:F1-score (Well-Defined): 78.4AUC (Well-Defined): 88.1F1-score (Ill-Defined): 90.4AUC (Ill-Defined): 97.1Hadilou et al.(2025) [94]Classify histopathological images of oral epithelial dysplasia (OED)ViT-B16Histopathological images218 histopathological slide images, processed into total 7542 patchesACC: 94/97in 3-class/4-class scenarioThomas et al.(2025) [95]Precise localization and classification of dental cystsDentViT-RCNNPanoramic radiographs137 imagesACC: 99.99 (Segmentation)98 (Classification)CancerFlügge et al.(2023) [96]Detection of oral squamous cell carcinoma (OSCC)Swin-TransformerClinical intraoral photographs1406 imagesACC: 98.58F1-score: 98.57SEN: 98.57SPE: 98.59Vinayahalingam et al.(2024) [97]Detect and classify the oral potentially malignant disorders (OPMD) including leukoplakia and oral OLP and to distinguish them from OSCCMask R-CNN with a Swin TransformerIntraoral photographs4,161 images with pixel-level annotationsOSCC:F1-score: 85.2AUC: 97.4OLP:F1-score: 82.5AUC: 94.8.Leukoplakia:F1-score: 79.6 AUC: 93.8.Authors (year)Aim of the StudyTransformer model usedDataset featureDataset sizePrimary PerformanceMetrics and Values (%)Anomalies/diseases detection, classification and localization (con’t)Cancer (con’t)Maia et al.(2024) [98]Investigate the effectiveness of AI models for classifying oral cancer, including OSCC, dysplasia, and non-dysplasia.Compare with ViT, PiT, CoaTHistopathological imagesP-NDB-UFES: 3763 images, with 1930 Dysplasia, 1126 OSCC, 707 Non-dysplasiaBest:CNN: DenseNet-121 (BCC: 91.91, PRE: 91.93)Transformer: CoaT (BCC: 90.84)FSL: ProtoNet (BCC:88.42)Song et al.(2024) [99]Binary classification (suspicious vs. non-suspicious) of oral cancer images captured using mobile devicesSwin-B, ViT-BMobile-based oral cavity imagesTotal 2434 Annotated imaginesSwin Transformer:ACC: 88.7SEN: 90.5SPE: 87.0Impacted teethKüçük et al.(2025) [100]Detect impacted teethHybrid of YOLOv8 and RT-DETR,Panoramic radiographs407 panoramic radiographsmAP50: 98.3F1-score: 96PRE: 93.8Recall: 99.2FluorosisXu et al.(2025) [101]Diagnosis of dental fluorosisMLTrMRDental images200 dental imagesACC: 80.19F1-score: 75.79qw Kappa: 81.28Anomalies/diseases detection, classification and localization (con’t)Fluorosis (con’t)Gu et al.(2024) [102]Automatic grading diagnosis of dental fluorosis by segmentation and classificationLKU-Net, FusionDentNetTooth images131 patients’ imagesSegmentation:ACC: 95.25PRE: 94.06Classification:ACC: 93.3PRE: 85.7Dental plaqueSong et al.(2024) [103]Dental plaque segmentationCenterFormerOral endoscope images2,884 imagesACC: 86.51mIoU: 75.86mPA: 85.94Tez et al.(2025) [104]Detecting and segmenting dental plaque in the young permanent teethUNet TransformerIntraoral images506 intraoral images of anterior young permanent teeth from 31 children (ages 8–13)IoU: 78.45Dice score: 82.15OthersWang et al.(2025) [105]Detection of dens evaginatusResNet-18Panoramic radiographs1410 panoramic radiographsSEN: 86.0SPE: 92.0F1-score: 88.66AUC: 95.60Dujic et al.(2023) [106]Detecting periodontal bone loss (PBL)ViT-baseViT-largeBEiT-baseBEiT-largeDeiT-basePeriapical dental radiographs21,819 periapical radiographsACC: 85.2 (ViT-base)AUC: 0.918 (ViT-base)SEN: 91.3 (BEiT-large)SPE: 74.5 (ViT-base)PPV: 89.1 (ViT-base)NPV: 77.5 (BEiT-large)Çelebi et al.(2024) [107]Automatically detect the boundaries of pathological conditions and infections in the maxillary sinusRes-Swin-UNetCBCT298 imagesIoU: 84.71F1-score: 91.72PRE: 91.57Recall: 91.87ACC: 99.87Anomalies/diseases detection, classification and localization (con’t)Others (con’t)Tang et al.(2024) [108]Automatically classify midpalatal suture (MPS) maturation stagesCNN-enhanced ViTCBCT2,518 labeled CBCT imagesACC: 95.75PRE: 95.15Recall: 93.15F1-score: 94.13Vinayahalingam et al.(2024) [109]Detect carotid artery calcifications (CAC)Faster R-CNN w/ Swin-TPanoramic radiographsTotal annotated 6,404 panoramic radiographsPRE: 89.5Recall: 88.1F1-score: 88.8SPE: 98.7Takano et al.(2025) [110]Improve the automatic detection of calcification regionsTransFusePanoramic radiographs578 annotated dental panoramic radiographs, split into left/right halves, resulting in 1,156 imagesDice: 63.41IoU: 50.43Lu et al.(2023) [111]Craniomaxillofacial (CMF) anatomical landmark localizationCMF-NetCBCT150 subjects and each with 24 manually labeled skeletal landmarksMRE: 1.108 mmSDR: 79.28Resul et al.(2025) [112]Automated periodontitis diagnosisAPD-FFNetPanoramic radiographs337 panoramic dental radiographsACC: 94F1-score: 93.88AUC: 93.47Lin et al.(2025) [113]Detect whether the mandibular third molar compresses the inferior alveolar nerve (IAN)Transformer-based ConvNeXt_v2Panoramic radiographs3,295 Panoramic radiographs images (after augmentation).Segmentation ACC: 97.92Chang et al.(2025) [114]Automated orthodontic diagnosisSPMA networkCephalograms3634 lateral cephalogramsACC: 90.02MR: 71.38Anomalies/diseases detection, classification and localization (con’t)Others (con’t)Zhang et al.(2025) [115]Evaluate the performance of the ViT for classifying molars with or without furcation involvement (FI)ViTPanoramic radiographsExtracted 1,568 tooth images from 506 radiographs, 7,840 images after augmentationPRE: 98Recall: 92F1-score: 95ACC: 92AUC: 87Benakatti et al.(2025) [116]Evaluate DETR for identifying different dental implant systems from radiographic imagesDETRRadiographic images1138 radiographs (868 periapical radiographs, 270 orthopantomograms)PRE: 83Recall: 89F1-score: 82Çelik et al.(2025) [117]Explore the impact of various data augmentation techniques on automated segmentation of dental restorations (implants, prostheses, and fillings)SegFormer, Swin-T Knet, PCPVT-L UperNet, SVT-S FPNPanoramic radiographs204 panoramic radiographs from the Tufts Dental DatabaseIoU: 62–82Dice score: 75–90ACC, Accuracy; AP: Average Precision; ASD, average surface distance; AUC, Area Under the ROC Curve; CBCT, Cone Beam Computed Tomography; clDice, centerline Sørensen–Dice coefficient; Dice, Sørensen–Dice coefficient; DSC, Dice Similarity Coefficient; HD, Hausdorff distance; IoU, Intersection over Union; JI, Jaccard Index; mACC, Mean Class Accuracy; mAP: mean Average Precision; mAR: Mean Average Recall; MR: Match Ratio; MRE: Mean Radial Error; mIoU, mean intersection over union; NPV: negative predictive value; PA: Pixel Accuracy; PPV, Positive Predictive Value; PRE, Precision; SDR: Successful Detection Rate; SEN, Sensitivity; SPE, Specificity; Characteristics of the included studies of the detection, classification and localization Gao et al. (2022) [31] ACC: 91.3 AUC: 99.7 Sifat et al. (2024) [68] For Dataset A/B: ACC: 90/100 Wang et al. (2024) [69] 37,828 pairs + 11 sequences of LiDAR-scanned outdoor scenes. On Shining3D: RRE: 0.12° On Oral3: RRE: 0.197° On KITTI: RRE: 0.23° Zhao et al. (2025) [70] ACC: 98.48 PRE: 97.29 Recall: 97.21 F1-score: 97.21 mAP: 98.12 Fernandes et al. (2025) [71] ACC: CNN: 0.71–0.99 ViT: 0.74–0.98 gMLP: 0.65–0.98 AUC: CNN: 0.77–1.00 ViT: 0.80–1.00 gMLP: 0.73–1.00 Liu et al. (2025) [72] mAP: 79.0 mAR: 65.9 Çelik et al. (2025) [73] ESRT + DenseNet (SF = 4): ACC: 92.3 F1-score 92.2 Felsch et al. (2023) [30] F1-score: 97.7 IoU: 95.9 AP: 97.7 ACC: 97.8 ACC: 85.57 PREC: 88.32 Recall: 83.17 F1-score: 85.67 AUC: 92.23 Jiang et al. (2021) [74] PRE: 62.3 Recall: 57.9 mAP: 56.9 FPS: 20.24 Priya et al. (2024) [75] ACC: 96.59 PRE: 96.55 Ying et al. (2024) [76] Compare with DETR, Trans-UNet F1-score YOLOv5: 87; DETR: 86; UNet: 82; Trans-UNet: 80 Youden Index: YOLOv5: 76; DETR: 73; UNet: 69; Trans-UNet: 64 Ying et al. (2022) [77] Dice: 74.87 PRE: 74.43 Neeraja et al. (2025) [78] ACC: 98.53 Dice: 95.93 IoU: 92.19 Chaves et al. (2024) [79] FROC AUC: 80.6/80.4 F1-score: 68.9/71.9 for primary and secondary caries detection, respectively Gao et al. (2024) [80] PRE: 64.1 Recall: 92.1 F1-score: 75.6 Huang et al. (2025) [81] Dice: 76.8 Jaccard: 64.4 Alsakar et al. (2024) [82] PRE: 95.7 SEN: 95.4 SPE: 95.7 ACC: 95.6 Nour et al. (2025) [83] PRE: 90 Recall: 79 mAP50: 74 ACC: 92 Xu et al. (2025) [84] Segmentation mAP: 84.9 Disease Detection: AP75: 64.5 Li et al. (2025) [85] Aljabar et al. (2024) [86] ACC: 96(ViT) 94 (AlexNet + RBF SVM + PCA) Asif et al. (2025) [87] ACC: 99.03 (MOD), 97.83(Oral Cancer), 94.23 (Histopathology) F1 ≈ 99 AUC: 99.98 Peng et al. (2025) [88] Liu et al. (2025) [89] ACC: 92.5 PRE: 90.2 Recall: 91.8 F1-score: 91.0 Chen et al. (2024) [90] Ahmed et al. (2023) [91] Panyarak et al. (2025) [92] Best performance (CrossViT-18 with ECAP): ACC: 72.6 Nistelrooij et al. (2025) [93] Mask R-CNN + Swin-Tiny: F1-score (Well-Defined): 78.4 AUC (Well-Defined): 88.1 F1-score (Ill-Defined): 90.4 AUC (Ill-Defined): 97.1 Hadilou et al. (2025) [94] ACC: 94/97 in 3-class/4-class scenario Thomas et al. (2025) [95] ACC: 99.99 (Segmentation) 98 (Classification) Flügge et al. (2023) [96] ACC: 98.58 F1-score: 98.57 SEN: 98.57 SPE: 98.59 Vinayahalingam et al. (2024) [97] OSCC: F1-score: 85.2 AUC: 97.4 OLP: F1-score: 82.5 AUC: 94.8. Leukoplakia: F1-score: 79.6 AUC: 93.8. Primary Performance Metrics and Values (%) Maia et al. (2024) [98] Best: CNN: DenseNet-121 (BCC: 91.91, PRE: 91.93) Transformer: CoaT (BCC: 90.84) FSL: ProtoNet (BCC:88.42) Song et al. (2024) [99] Swin Transformer: ACC: 88.7 SEN: 90.5 SPE: 87.0 Küçük et al. (2025) [100] mAP50: 98.3 F1-score: 96 PRE: 93.8 Recall: 99.2 Xu et al. (2025) [101] ACC: 80.19 F1-score: 75.79 qw Kappa: 81.28 Gu et al. (2024) [102] Segmentation: ACC: 95.25 PRE: 94.06 Classification: ACC: 93.3 PRE: 85.7 Song et al. (2024) [103] ACC: 86.51 mIoU: 75.86 mPA: 85.94 Tez et al. (2025) [104] IoU: 78.45 Dice score: 82.15 Wang et al. (2025) [105] SEN: 86.0 SPE: 92.0 F1-score: 88.66 AUC: 95.60 Dujic et al. (2023) [106] ViT-base ViT-large BEiT-base BEiT-large DeiT-base ACC: 85.2 (ViT-base) AUC: 0.918 (ViT-base) SEN: 91.3 (BEiT-large) SPE: 74.5 (ViT-base) PPV: 89.1 (ViT-base) NPV: 77.5 (BEiT-large) Çelebi et al. (2024) [107] IoU: 84.71 F1-score: 91.72 PRE: 91.57 Recall: 91.87 ACC: 99.87 Tang et al. (2024) [108] ACC: 95.75 PRE: 95.15 Recall: 93.15 F1-score: 94.13 Vinayahalingam et al. (2024) [109] PRE: 89.5 Recall: 88.1 F1-score: 88.8 SPE: 98.7 Takano et al. (2025) [110] Dice: 63.41 IoU: 50.43 Lu et al. (2023) [111] MRE: 1.108 mm SDR: 79.28 Resul et al. (2025) [112] ACC: 94 F1-score: 93.88 AUC: 93.47 Lin et al. (2025) [113] Chang et al. (2025) [114] ACC: 90.02 MR: 71.38 Zhang et al. (2025) [115] PRE: 98 Recall: 92 F1-score: 95 ACC: 92 AUC: 87 Benakatti et al. (2025) [116] PRE: 83 Recall: 89 F1-score: 82 Çelik et al. (2025) [117] IoU: 62–82 Dice score: 75–90 ACC, Accuracy; AP: Average Precision; ASD, average surface distance; AUC, Area Under the ROC Curve; CBCT, Cone Beam Computed Tomography; clDice, centerline Sørensen–Dice coefficient; Dice, Sørensen–Dice coefficient; DSC, Dice Similarity Coefficient; HD, Hausdorff distance; IoU, Intersection over Union; JI, Jaccard Index; mACC, Mean Class Accuracy; mAP: mean Average Precision; mAR: Mean Average Recall; MR: Match Ratio; MRE: Mean Radial Error; mIoU, mean intersection over union; NPV: negative predictive value; PA: Pixel Accuracy; PPV, Positive Predictive Value; PRE, Precision; SDR: Successful Detection Rate; SEN, Sensitivity; SPE, Specificity; Classification often follows segmentation, with semantic segmentation categorizing teeth anatomically and instance segmentation distinguishing individual teeth for anomaly detection. Gao et al. proposed a Grouped Bottleneck Transformer integrating 3D convolutions, self-attention, and bottlenecks, improving CBCT-based classification with reduced computational cost [31]. For prosthodontics, Sifat et al. combined ViTs with 3D CNNs and Explainable AI, achieving 90–100% accuracy on 3D dental scans [68]. Wang et al. introduced GeoSwin, a geometric multiscale Transformer that enhanced orthodontic point cloud registration [69]. Zhao et al. integrated a RoI Transformer with ResNet-50, improving precision by aligning detection with natural tooth orientation [70]. Fernandes et al. showed ViT and gMLP achieved performance comparable to or superior to CNNs in radiographic classification [71]. Liu et al. developed a segmentation-free two-stage method using geodesic heatmap encoding and PTv3 backbone for landmark detection, outperforming prior approaches [72]. Çelik et al. applied ESPCN-based super-resolution to radiographs, improving image clarity and classification accuracy [73]. Once teeth are segmented and classified, Transformer-based systems extend to pathology detection. Caries was most frequently studied (n = 17), alongside oral lesions (n = 9), cancers (n = 5), impacted teeth (n = 5), implants (n = 5), plaque (n = 3), fluorosis (n = 2), and other conditions (n = 15), underscoring their versatility. For dental caries, Transformers have been applied across endoscopic images, intraoral photos, bitewing, periapical, and panoramic radiographs. Felsch et al. used SegFormer for lesion and MIH detection in photos with high precision [30]. Zhou et al. developed a Swin Transformer leveraging tooth type–specific features, achieving near expert-level pediatric caries diagnosis [32]. Jiang et al. proposed RDFNet, a Transformer-augmented YOLOv5s with FReLU activation for efficient digital caries detection [74]. Priya et al. combined ATDUnet++ with ViT-MRDGRU, achieving 96.59% accuracy [75]. Ying et al. designed a Transformer-enhanced U-Net for automated X-ray caries segmentation [77], while Neeraja et al. introduced DentC_BSFLNet (Swin-UNETI + Involution), improving panoramic caries detection accuracy [78]. Chaves et al. employed Mask R-CNN with Swin backbone for bitewing caries [79]. Gao et al. developed GFDet, a PVT-based model with feature fusion and adaptive label assignment for endoscopic caries detection [80]. Huang et al. proposed AEDD-Net, a CNN–Transformer hybrid with dual decoders and edge-aware loss, achieving state-of-the-art X-ray segmentation [81]. In a benchmark comparison, Ying et al. found YOLOv5 outperformed DETR, UNet, and TransUNet on periapical caries, followed by TransUNet [76]. Several studies tackled multiple conditions. Alsakar et al. built a dual-stream MobileNetV2 + Swin Transformer ensemble for panoramic multi-label classification [82]. Nour et al. integrated YOLOv8 and RT-DETR for high-accuracy detection of caries, implants, fillings, missing teeth, lesions, and root treatments [83]. Xu et al. proposed UCL-Net, a lightweight CNN with cross-scale attention and uncertainty-aware learning, excelling in caries, deep caries, lesions, and impactions [84]. Li et al. developed LABANet, a Swin-based segmentation network with synthetic data balancing, accurately detecting wisdom teeth, caries, and residual roots [85]. Alabd-Aljabar et al. achieved 96% accuracy with ViT for multi-condition diagnosis, surpassing CNN–SVM approaches [86]. Asif et al. introduced OralTransNet, a lightweight MobileNet–Transformer hybrid, exceeding 99% diagnostic accuracy with low computational cost [87]. Peng et al. proposed Oralformer, a parallel Transformer segmentation model with uncertainty-driven loss, achieving state-of-the-art accuracy on the ODS dataset [88]. Liu et al. infused agent attention into a Conformer, significantly boosting disease classification accuracy [89]. Transformer models extended to periapical and jaw lesions across OCT, panoramic, histopathology, and CBCT. Chen et al. showed that pretrained Swin-UNETR outperformed U-Net in low-data CBCT lesion segmentation [90]. Ahmed et al. introduced TED, a ViT-based framework improving OCT and CT lesion detection under noisy conditions [91]. Panyarak et al. demonstrated CrossViT models outperformed ResNet/ViT in classifying four jaw lesion types using ECAP augmentation [92]. Hadilou et al. achieved up to 97% accuracy for oral epithelial dysplasia grading with ViTs, surpassing CNNs [94]. Van Nistelrooij et al. found Transformers superior in classifying osteolytic lesions [93]. Thomas et al. proposed DentViT-RCNN, combining Mask R-CNN, ViT, and GAN augmentation for high-accuracy cyst diagnosis [95]. For oral cancer, Flügge et al. applied Swin Transformer to clinical photos, achieving 98.6% accuracy and AUC = 0.99 for oral squamous cell carcinoma [96]. They later used ViT-based Mask R-CNN for carcinoma, oral lichen planus, and leukoplakia with strong results [97]. Song et al. found Swin Transformer outperformed ViT and CNNs on mobile oral cancer images [99]. Maia et al. confirmed comparable CNN and Transformer performance in histopathology using a curated dataset [98]. Other anomalies included in the last five years’ studies:Impacted teeth: Küçük et al. combined YOLOv8, RT-DETR, GFP-GAN, and Weighted Boxes Fusion, achieving 98.3% mAP [100].Implants: Studied within multi-anomaly frameworks [82, 83, 86].Plaque: Song et al. introduced CenterFormer for cluster-enhanced detection [103]; Tez et al. applied a Transformer-UNet surpassing pediatric dentists in plaque detection [104].Fluorosis: Gu et al. proposed FusionDentNet, a U-Net + CNN–Transformer pipeline for grading severity [102]; Xu et al. developed MLTrMR with adaptive masking, achieving state-of-the-art DFID results [101].Other pathologies: Transformers have been used for dens evaginatus [105], periodontal bone loss [106], maxillary sinus disease [107], midpalatal suture maturation [108], carotid artery calcification [109, 110], craniomaxillofacial landmark localization [111], periodontitis [112], inferior alveolar nerve compression [113], orthodontic malocclusion [114], furcation involvement [115], and implant type classification [116, 117]. Impacted teeth: Küçük et al. combined YOLOv8, RT-DETR, GFP-GAN, and Weighted Boxes Fusion, achieving 98.3% mAP [100]. Impacted teeth: Küçük et al. combined YOLOv8, RT-DETR, GFP-GAN, and Weighted Boxes Fusion, achieving 98.3% mAP [100]. Implants: Studied within multi-anomaly frameworks [82, 83, 86]. Implants: Studied within multi-anomaly frameworks [82, 83, 86]. Plaque: Song et al. introduced CenterFormer for cluster-enhanced detection [103]; Tez et al. applied a Transformer-UNet surpassing pediatric dentists in plaque detection [104]. Plaque: Song et al. introduced CenterFormer for cluster-enhanced detection [103]; Tez et al. applied a Transformer-UNet surpassing pediatric dentists in plaque detection [104]. Fluorosis: Gu et al. proposed FusionDentNet, a U-Net + CNN–Transformer pipeline for grading severity [102]; Xu et al. developed MLTrMR with adaptive masking, achieving state-of-the-art DFID results [101]. Fluorosis: Gu et al. proposed FusionDentNet, a U-Net + CNN–Transformer pipeline for grading severity [102]; Xu et al. developed MLTrMR with adaptive masking, achieving state-of-the-art DFID results [101]. Other pathologies: Transformers have been used for dens evaginatus [105], periodontal bone loss [106], maxillary sinus disease [107], midpalatal suture maturation [108], carotid artery calcification [109, 110], craniomaxillofacial landmark localization [111], periodontitis [112], inferior alveolar nerve compression [113], orthodontic malocclusion [114], furcation involvement [115], and implant type classification [116, 117]. Other pathologies: Transformers have been used for dens evaginatus [105], periodontal bone loss [106], maxillary sinus disease [107], midpalatal suture maturation [108], carotid artery calcification [109, 110], craniomaxillofacial landmark localization [111], periodontitis [112], inferior alveolar nerve compression [113], orthodontic malocclusion [114], furcation involvement [115], and implant type classification [116, 117]. Clinical generative applications A total of 25 studies applied Transformer-based models for dental anomaly diagnosis (Table 3). Most generated clinically relevant results, underscoring the applicability of Transformer-based AI in dental practice. Seventeen focused on generative clinical applications, grouped into seven domains:Dental Implant Position Prediction: Yang et al. developed ImplantFormer, a ViT–ResNet hybrid predicting implant positions from CBCT crown images by leveraging neighboring tooth texture, outperforming CNN detectors [118]. The same group later proposed TSIPR, a two-stream YOLOv7-X + ViT model, which surpassed existing CNN and Transformer baselines in implant localization [119].Dental Age and Sex Estimation: Fan et al. introduced DASE-Net, a semi-supervised CNN–Transformer hybrid achieving MAE of 2.61 years for age and 95.5% accuracy for sex classification on panoramic radiographs [120]. Arian et al. proposed PENViT, an EfficientNet–ViT hybrid providing explainable, state-of-the-art predictions validated by FullGrad, outperforming CNNs and triplet networks [121].Restorative Dentistry − 3D Structure Generation: Several models focused on 3D restoration. Shen et al. developed TranSDFNet for realistic RPD clasp generation using signed distance fields, outperforming prior models [122]. Saleh et al. showed ShapeFormer- and MeshDiffusion-based methods could reconstruct missing teeth from limited 3D data [123]. Hosseinimanesh et al. proposed a point-to-mesh Transformer generating crowns with high precision via Chamfer loss and margin constraints [124]. Alsheghri et al. integrated uncertainty quantification for margin line generation, achieving a 0.137 mm Chamfer Distance [125]. Liu et al. designed TUCNet, combining channel/spatial attention with Transformer blocks for point cloud upsampling [126], and later introduced DDMF, a CBCT–IOS fusion pipeline for accurate crown–root–bone reconstruction [127]. Yang et al. presented DCrownFormer+, a morphology-aware Transformer enhancing occlusal detail and surface smoothness in crown mesh generation [128]. Zhang et al. developed MPGLNet, a progressive fusion network restoring global structures and fine details in point cloud completion on the FID-Comp3D dataset [129].Therapy Outcome Prediction: Lin et al. introduced TransFArchNet, a multi-scale Transformer predicting dental arch curves from facial point clouds with sub-millimeter accuracy [130]. Trachoo et al. proposed a three-phase ResNet101V2–RetinaNet–ViT pipeline for detecting and assessing impacted third molars [131]. Lee et al. applied a ViT for third molar detection and IAN injury risk prediction [132]. Li et al. created AGMB-Transformer, achieving 90.2% accuracy in root canal therapy evaluation by integrating global and local features [133]. Ma et al. proposed F-VAMP, a hierarchical Transformer-based model that improved tooth position prediction with adaptive geometry and physiological optimization [134].Dental Dataset Enhancement and Denoising: Ahmed et al. developed TEAR, a Transformer–Autoencoder hybrid for OCT denoising, improving visual quality and classification [135]. Sivaanpu et al. proposed HCTSpeckle, a CNN–Transformer model for ultrasound denoising and segmentation with high efficiency [136]. Li et al. introduced Semi-SFTrans, a semi-supervised spatial-frequency Transformer reducing metal artifacts in maxillofacial CT [137]. Yu et al. presented ToothMaker, a diffusion framework generating realistic panoramic radiographs with disentangled style/layout, boosting segmentation and VQA [138]. Ha et al. designed TransUNet-SegGAN to generate high-quality T2-weighted MR images from PD-weighted scans, reducing scan time for TMD patients [139].Unsupervised Learning and Equipment Impact: Farook et al. showed scanner and operator variability minimally affect scan accuracy, but controlled volumetric augmentation was key for unsupervised learning. Their 3D ViT achieved 96.2% classification accuracy [140].Predict inflammasome peptide sequences: Yadalam et al. combined Transformer and Graph Attention Networks to predict AIM-2 inflammasome peptide sequences in periodontal inflammation, outperforming Recurrent Neural Network(RNN)-CNN and highlighting potential for targeted therapeutics [141].Table 3Characteristics of the clinical generative applicationsAuthors (year)Aim of the StudyTransformer model usedDataset featureDataset sizePrimary Performance Metrics and Values (%)Prediction & Other Clinical ApplicationsYang et al.(2024) [118]Regress dental implant positions from CBCT imagesImplantFormerCBCTSelected 3045 annotated 2D slices of tooth crownAP: 13.7F1-score: 13.6Fan et al.(2023) [120]Estimate both dental age and biological sexDASE-NetOrthopantomograms (OPGs) imagesTotal 15,295 OPGsMAE: 2.61 (years)AUC: 95.54Li et al.(2022) [133]Evaluation of root canal therapy resultsAGMB-TransformerPeriapical radiographs245 X-ray imagesACC: 90.20AUC: 95.63SEN: 91.39SPE: 95.09F1-score: 90.48Shen et al.(2023) [122]Generate removable partial denture (RPD) claspsTranSDFNetIntraoral 3D dental scans39 samplesF-score@1%: 82.88CD-L2: 0.00122EMD: 0.0713Lin et al.(2025) [130]Predict personalized dental arch curvesTransFArchNetCBCT467 CBCT scansHD90 (XY-axis): 3.50 mmHD90 (XYZ-axis): 4.06 mmSaleh et al.(2024) [123]Reconstructing single and multiple missing teethTransformer-based models (TM) and diffusion-based models (DM)Intraoral 3D dental scans2064 3D jaw modelTM highest accuracy:Tooth 34 (RMS = 0.21)Farook et al.(2023) [140]Investigate how intraoral scanners, operators, and data augmentation affect unsupervised deep learning3DViTIntraoral scan22 dental cast sets. Total of 110 native scans plus 68 synthetic augmented scansACC: 96.2 (in the augmented dataset)Yang et al.(2024) [119]Improve the accuracy of dental implant position predictionTSIPRCBCTLabeled 3045 slicesAP75: 15.7F1 Score: 15.6Prediction & Other Clinical Applications (con’t)Hosseinimanesh et al.(2025) [124]Automatically generates personalized 3D dental crown meshesTransformer-based point cloud encoder–decoderIntraoral 3D dental scans point clouds557 in totalCD-L1: 54.39Ahmed et al.(2024) [135]Remove speckle and other noise artifacts from optical coherence tomography (OCT) imagesTEAROCT images16 noisy/clean OCT image pairs + 10 healthy/decayed samples + 84,484 noisy images from 4 classesF1-score: improved from 87 to 92Arian et al.(2023) [121]Estimating age and biological genderPENViTPanoramic radiographsTotal 2,630 labeled samples.ACC: 70.54Liu et al.(2023) [127]3D reconstruction of crown-root-bone structures with multimodal fusion of CBCT and intraoral scan (IOS)DDMFFusing CBCT and IOS dataCBCT Data: 503 scans with 9,651 manually annotated slicesIOS Meshes: 28,559 scansHD: 0.47 mmTrachoo et al.(2025) [131]Predict the difficulty level of surgical removal of impacted mandibular third molars (LM3)ViTPanoramic radiographs1367 LM3 casesClassification accuracy: 86.71(ResNet101V2)Detection average precision: 99.28(RetinaNet)Predicting removal difficulty levels accuracy: 78.99(Vision Transformer)Sivaanpu et al.(2024) [136]Reducing speckle noise in ultrasound (US) imagesHCTSpeckleUS imaging4,578 imagesMSE: as low as 13.04PSNR: up to 37.85 dBSSIM: up to 0.98Prediction & Other Clinical Applications (con’t)Lee et al.(2022) [132]Predicts both the extraction difficulty and the likelihood of inferior alveolar nerve (IAN) injury for mandibular third molarsViT-L/32Panoramic radiographs5397 annotated panoramic radiographsDetection:mAP@0.5: 99.0Extraction Difficulty/IAN injury Prediction:ACC: 83.5/81.1Alsheghri et al.(2024) [125]Automatically generates dental margin lines for the crownAdaPoinTrIntraoral scans913 samplesCD: 0.137 mmHD: 0.242 mmLiu et al.(2023) [126]Complete and upsample sparse and incomplete 3D dental point cloudsTUCNetIntraoral 3D dental scans point cloudsPCN Dataset; Teeth Dataset (FDDI): elf-constructed dataset from 132 dental cases, generating ~3000 partial-complete point cloud pairsCD-L1 (×10− 3): 222CD-L2 (×10− 3): 194EMD: 775P2S: 0.0112Li et al.(2025) [137]Metal artifact reduction (MAR) in maxillofacial CT imagesSemi-SFTransCT imagesSimulated: 28,036 paired images (45 cases).Clinical: 58 artifact cases (1,087 slices); 22 clinical cases with intraoral scans.Simulated:PSNR = 46.1 dB,SSIM = 0.9913Clinical:Likert = 4.46,Dice = 86.3,HD = 11.3Yu et al.(2025) [138]Generating realistic panoramic dental radiographsStable Diffusion 1.5Panoramic radiographsDental X-ray images: 3,657 Segmentation masks: 3,441 annotated imagesDownstream tasks: 2,900 annotated panoramic dental images and 18,240 VQA pairsFID (↓ better): 72.91IS (↑ better): 2.85MS-SSIM (↑ better): 0.63Prediction & Other Clinical Applications (con’t)Yang et al.(2025) [128]Automatically generate high-fidelity, realistic dental crown prosthesis meshes from 3D scans of preparation and antagonist teethDCrownFormer+3D scans3,144 scansCD: 41.004F1-score: 93.1HD: 62.932Yadalam et al.(2025) [141]Predict AIM-2 inflammasome peptide sequences associated with periodontal inflammationTransformerFASTA sequencesUniProt databaseACC: 84SEN: 83AUC: 89.1MCC: 68Zhang et al.(2025) [129]Generate high-resolution, complete, and dense dental point cloudsMPGLNet3D scans16,384 points for complete clouds, 2,048 points for partialCDL1:199.038CDL2: 94.218Ha et al.(2025) [139]Generate T2-weighted (T2w) MR images from proton density-weighted (PDw) images in temporomandibular disorder (TMD) patientsTransUNet-SegGANMRI scans7,226 MR slicesSSIM: 82.28FID: 23.85Ma et al.(2025) [134]Tooth position predictionF-VAMPDental position prediction dataset834 Dental position prediction cases constructed from pathological cases.CSA: 87.67MEt: 0.8338 mmSilva et al.(2024) [142]Evaluate the performance in describing radiolucent lesions and in establishing diagnosesChatGPT-3.5Panoramic radiographs28 casesDescriptive Accuracy Overall Mean Score: 0.56 ± 0.48Differential Diagnosis Accuracy: Rank-1: 25%Rank-2: 57.14%Rank-3: 67.86%Selected Generative Pre-trained Transformer (GPT) Clinical ApplicationsJi et al.(2025) [143]Evaluate the capability of GPT-4 in answering open-ended questions related to oral cancerGPT-4Open-ended questions60 selected from a book titled “100 Questions about Oral Cancer Health KnowledgeEvaluated using a modified Global Quality Scale [1–5]: Most answers were scored at 4 or 5Salmanpour et al.(2025) [144]Evaluate performance in determining the labiolingual localization of maxillary impacted canines and detecting resorption in adjacent incisorsChatGPT-4Panoramic radiographs & CBCT105 patients for localization task; 50 patients for resorption detectionLocalization:ACC: 37.1Resorption Detection:ACC: 46.0Bordin et al.(2025) [145]Evaluate the accuracy and reliability of ChatGPT-3.5 in answering questions related to dental traumaChatGPT 3.5Open-ended questions45 questions generated by ChatGPT-3.5, evaluated by five experienced endodontistsMean scores:Avulsion: 4.40 ± 0.717Intrusion: 4.13 ± 0.794General Dental Trauma: 3.97 ± 0.885Chuang et al.(2024) [146]Evaluate the effectiveness of using GPT-J for prompt generation to create seed data for training RoBERTa-based named entity recognition (NER) modelsGPT-J-6BClinical notes from EDRs8,125 clinical notes focus on periodontitisF1 Score: 0.92–0.97Ameli et al.(2024) [147]Developing a Clinical Decision Support System (CDSS) that classifies the stage of the disease.BERTClinical notes from EDRs309 anonymized patient chartsStaging:ACC: 66Grading:ACC: 72Selected Generative Pre-trained Transformer (GPT) Clinical Applications (con’t)Molena et al.(2024) [148]Evaluate the accuracy, completeness, and reliability of responses generated by ChatGPT 3.5 in dentistryChatGPT 3.5Expert-formulated dental questions30 questionsAccuracy Scores [1–6]: 4.17Completeness Scores [1–3]: 2.07Giannakopoulos et al.(2023) [149]Evaluate and compare the performance of four LLMs in answering dentistry questionsChatGPT-3.5, ChatGPT-4Open-ended, evidence-based clinical dentistry question20 questionsAverage scores [1–10]:ChatGPT-4: 7.2ChatGPT-3.5: 5.9Google Bard: 5.7Microsoft Bing Chat: 5.4ACC, Accuracy; AP: Average Precision; AUC, Area Under the ROC Curve; CBCT, Cone Beam Computed Tomography; CD: Chamfer Distance; CD-L: Chamfer-L Distance; CSA: Correct Spatial Alignment; Dice coefficient; Dice, Sørensen Dice coefficient; DSC, Dice Similarity Coefficient; EMD: Earth Mover’s Distance; FID: Frechet Inception Distance; HD, Hausdorff distance; IoU, Intersection over Union; IS: Inception Score; MAE: mean absolute error; MCC: Matthews correlation coefficient; MEt: Mean; Translation Error; mAP: mean Average Precision; MSE: mean squared error; MS-SSIM: Multi-Scale Structural; Similarity Index Metric; mIoU, mean intersection over union; PA: Pixel Accuracy; PRE, Precision; PSNR: peak signal-to-noise ratio; RMS: Root mean Square; SEN, Sensitivity; SPE, Specificity; SSIM: structural similarity index Dental Implant Position Prediction: Yang et al. developed ImplantFormer, a ViT–ResNet hybrid predicting implant positions from CBCT crown images by leveraging neighboring tooth texture, outperforming CNN detectors [118]. The same group later proposed TSIPR, a two-stream YOLOv7-X + ViT model, which surpassed existing CNN and Transformer baselines in implant localization [119]. Dental Implant Position Prediction: Yang et al. developed ImplantFormer, a ViT–ResNet hybrid predicting implant positions from CBCT crown images by leveraging neighboring tooth texture, outperforming CNN detectors [118]. The same group later proposed TSIPR, a two-stream YOLOv7-X + ViT model, which surpassed existing CNN and Transformer baselines in implant localization [119]. Dental Age and Sex Estimation: Fan et al. introduced DASE-Net, a semi-supervised CNN–Transformer hybrid achieving MAE of 2.61 years for age and 95.5% accuracy for sex classification on panoramic radiographs [120]. Arian et al. proposed PENViT, an EfficientNet–ViT hybrid providing explainable, state-of-the-art predictions validated by FullGrad, outperforming CNNs and triplet networks [121]. Dental Age and Sex Estimation: Fan et al. introduced DASE-Net, a semi-supervised CNN–Transformer hybrid achieving MAE of 2.61 years for age and 95.5% accuracy for sex classification on panoramic radiographs [120]. Arian et al. proposed PENViT, an EfficientNet–ViT hybrid providing explainable, state-of-the-art predictions validated by FullGrad, outperforming CNNs and triplet networks [121]. Restorative Dentistry − 3D Structure Generation: Several models focused on 3D restoration. Shen et al. developed TranSDFNet for realistic RPD clasp generation using signed distance fields, outperforming prior models [122]. Saleh et al. showed ShapeFormer- and MeshDiffusion-based methods could reconstruct missing teeth from limited 3D data [123]. Hosseinimanesh et al. proposed a point-to-mesh Transformer generating crowns with high precision via Chamfer loss and margin constraints [124]. Alsheghri et al. integrated uncertainty quantification for margin line generation, achieving a 0.137 mm Chamfer Distance [125]. Liu et al. designed TUCNet, combining channel/spatial attention with Transformer blocks for point cloud upsampling [126], and later introduced DDMF, a CBCT–IOS fusion pipeline for accurate crown–root–bone reconstruction [127]. Yang et al. presented DCrownFormer+, a morphology-aware Transformer enhancing occlusal detail and surface smoothness in crown mesh generation [128]. Zhang et al. developed MPGLNet, a progressive fusion network restoring global structures and fine details in point cloud completion on the FID-Comp3D dataset [129]. Restorative Dentistry − 3D Structure Generation: Several models focused on 3D restoration. Shen et al. developed TranSDFNet for realistic RPD clasp generation using signed distance fields, outperforming prior models [122]. Saleh et al. showed ShapeFormer- and MeshDiffusion-based methods could reconstruct missing teeth from limited 3D data [123]. Hosseinimanesh et al. proposed a point-to-mesh Transformer generating crowns with high precision via Chamfer loss and margin constraints [124]. Alsheghri et al. integrated uncertainty quantification for margin line generation, achieving a 0.137 mm Chamfer Distance [125]. Liu et al. designed TUCNet, combining channel/spatial attention with Transformer blocks for point cloud upsampling [126], and later introduced DDMF, a CBCT–IOS fusion pipeline for accurate crown–root–bone reconstruction [127]. Yang et al. presented DCrownFormer+, a morphology-aware Transformer enhancing occlusal detail and surface smoothness in crown mesh generation [128]. Zhang et al. developed MPGLNet, a progressive fusion network restoring global structures and fine details in point cloud completion on the FID-Comp3D dataset [129]. Therapy Outcome Prediction: Lin et al. introduced TransFArchNet, a multi-scale Transformer predicting dental arch curves from facial point clouds with sub-millimeter accuracy [130]. Trachoo et al. proposed a three-phase ResNet101V2–RetinaNet–ViT pipeline for detecting and assessing impacted third molars [131]. Lee et al. applied a ViT for third molar detection and IAN injury risk prediction [132]. Li et al. created AGMB-Transformer, achieving 90.2% accuracy in root canal therapy evaluation by integrating global and local features [133]. Ma et al. proposed F-VAMP, a hierarchical Transformer-based model that improved tooth position prediction with adaptive geometry and physiological optimization [134]. Therapy Outcome Prediction: Lin et al. introduced TransFArchNet, a multi-scale Transformer predicting dental arch curves from facial point clouds with sub-millimeter accuracy [130]. Trachoo et al. proposed a three-phase ResNet101V2–RetinaNet–ViT pipeline for detecting and assessing impacted third molars [131]. Lee et al. applied a ViT for third molar detection and IAN injury risk prediction [132]. Li et al. created AGMB-Transformer, achieving 90.2% accuracy in root canal therapy evaluation by integrating global and local features [133]. Ma et al. proposed F-VAMP, a hierarchical Transformer-based model that improved tooth position prediction with adaptive geometry and physiological optimization [134]. Dental Dataset Enhancement and Denoising: Ahmed et al. developed TEAR, a Transformer–Autoencoder hybrid for OCT denoising, improving visual quality and classification [135]. Sivaanpu et al. proposed HCTSpeckle, a CNN–Transformer model for ultrasound denoising and segmentation with high efficiency [136]. Li et al. introduced Semi-SFTrans, a semi-supervised spatial-frequency Transformer reducing metal artifacts in maxillofacial CT [137]. Yu et al. presented ToothMaker, a diffusion framework generating realistic panoramic radiographs with disentangled style/layout, boosting segmentation and VQA [138]. Ha et al. designed TransUNet-SegGAN to generate high-quality T2-weighted MR images from PD-weighted scans, reducing scan time for TMD patients [139]. Dental Dataset Enhancement and Denoising: Ahmed et al. developed TEAR, a Transformer–Autoencoder hybrid for OCT denoising, improving visual quality and classification [135]. Sivaanpu et al. proposed HCTSpeckle, a CNN–Transformer model for ultrasound denoising and segmentation with high efficiency [136]. Li et al. introduced Semi-SFTrans, a semi-supervised spatial-frequency Transformer reducing metal artifacts in maxillofacial CT [137]. Yu et al. presented ToothMaker, a diffusion framework generating realistic panoramic radiographs with disentangled style/layout, boosting segmentation and VQA [138]. Ha et al. designed TransUNet-SegGAN to generate high-quality T2-weighted MR images from PD-weighted scans, reducing scan time for TMD patients [139]. Unsupervised Learning and Equipment Impact: Farook et al. showed scanner and operator variability minimally affect scan accuracy, but controlled volumetric augmentation was key for unsupervised learning. Their 3D ViT achieved 96.2% classification accuracy [140]. Unsupervised Learning and Equipment Impact: Farook et al. showed scanner and operator variability minimally affect scan accuracy, but controlled volumetric augmentation was key for unsupervised learning. Their 3D ViT achieved 96.2% classification accuracy [140]. Predict inflammasome peptide sequences: Yadalam et al. combined Transformer and Graph Attention Networks to predict AIM-2 inflammasome peptide sequences in periodontal inflammation, outperforming Recurrent Neural Network(RNN)-CNN and highlighting potential for targeted therapeutics [141]. Predict inflammasome peptide sequences: Yadalam et al. combined Transformer and Graph Attention Networks to predict AIM-2 inflammasome peptide sequences in periodontal inflammation, outperforming Recurrent Neural Network(RNN)-CNN and highlighting potential for targeted therapeutics [141]. Table 3Characteristics of the clinical generative applicationsAuthors (year)Aim of the StudyTransformer model usedDataset featureDataset sizePrimary Performance Metrics and Values (%)Prediction & Other Clinical ApplicationsYang et al.(2024) [118]Regress dental implant positions from CBCT imagesImplantFormerCBCTSelected 3045 annotated 2D slices of tooth crownAP: 13.7F1-score: 13.6Fan et al.(2023) [120]Estimate both dental age and biological sexDASE-NetOrthopantomograms (OPGs) imagesTotal 15,295 OPGsMAE: 2.61 (years)AUC: 95.54Li et al.(2022) [133]Evaluation of root canal therapy resultsAGMB-TransformerPeriapical radiographs245 X-ray imagesACC: 90.20AUC: 95.63SEN: 91.39SPE: 95.09F1-score: 90.48Shen et al.(2023) [122]Generate removable partial denture (RPD) claspsTranSDFNetIntraoral 3D dental scans39 samplesF-score@1%: 82.88CD-L2: 0.00122EMD: 0.0713Lin et al.(2025) [130]Predict personalized dental arch curvesTransFArchNetCBCT467 CBCT scansHD90 (XY-axis): 3.50 mmHD90 (XYZ-axis): 4.06 mmSaleh et al.(2024) [123]Reconstructing single and multiple missing teethTransformer-based models (TM) and diffusion-based models (DM)Intraoral 3D dental scans2064 3D jaw modelTM highest accuracy:Tooth 34 (RMS = 0.21)Farook et al.(2023) [140]Investigate how intraoral scanners, operators, and data augmentation affect unsupervised deep learning3DViTIntraoral scan22 dental cast sets. Total of 110 native scans plus 68 synthetic augmented scansACC: 96.2 (in the augmented dataset)Yang et al.(2024) [119]Improve the accuracy of dental implant position predictionTSIPRCBCTLabeled 3045 slicesAP75: 15.7F1 Score: 15.6Prediction & Other Clinical Applications (con’t)Hosseinimanesh et al.(2025) [124]Automatically generates personalized 3D dental crown meshesTransformer-based point cloud encoder–decoderIntraoral 3D dental scans point clouds557 in totalCD-L1: 54.39Ahmed et al.(2024) [135]Remove speckle and other noise artifacts from optical coherence tomography (OCT) imagesTEAROCT images16 noisy/clean OCT image pairs + 10 healthy/decayed samples + 84,484 noisy images from 4 classesF1-score: improved from 87 to 92Arian et al.(2023) [121]Estimating age and biological genderPENViTPanoramic radiographsTotal 2,630 labeled samples.ACC: 70.54Liu et al.(2023) [127]3D reconstruction of crown-root-bone structures with multimodal fusion of CBCT and intraoral scan (IOS)DDMFFusing CBCT and IOS dataCBCT Data: 503 scans with 9,651 manually annotated slicesIOS Meshes: 28,559 scansHD: 0.47 mmTrachoo et al.(2025) [131]Predict the difficulty level of surgical removal of impacted mandibular third molars (LM3)ViTPanoramic radiographs1367 LM3 casesClassification accuracy: 86.71(ResNet101V2)Detection average precision: 99.28(RetinaNet)Predicting removal difficulty levels accuracy: 78.99(Vision Transformer)Sivaanpu et al.(2024) [136]Reducing speckle noise in ultrasound (US) imagesHCTSpeckleUS imaging4,578 imagesMSE: as low as 13.04PSNR: up to 37.85 dBSSIM: up to 0.98Prediction & Other Clinical Applications (con’t)Lee et al.(2022) [132]Predicts both the extraction difficulty and the likelihood of inferior alveolar nerve (IAN) injury for mandibular third molarsViT-L/32Panoramic radiographs5397 annotated panoramic radiographsDetection:mAP@0.5: 99.0Extraction Difficulty/IAN injury Prediction:ACC: 83.5/81.1Alsheghri et al.(2024) [125]Automatically generates dental margin lines for the crownAdaPoinTrIntraoral scans913 samplesCD: 0.137 mmHD: 0.242 mmLiu et al.(2023) [126]Complete and upsample sparse and incomplete 3D dental point cloudsTUCNetIntraoral 3D dental scans point cloudsPCN Dataset; Teeth Dataset (FDDI): elf-constructed dataset from 132 dental cases, generating ~3000 partial-complete point cloud pairsCD-L1 (×10− 3): 222CD-L2 (×10− 3): 194EMD: 775P2S: 0.0112Li et al.(2025) [137]Metal artifact reduction (MAR) in maxillofacial CT imagesSemi-SFTransCT imagesSimulated: 28,036 paired images (45 cases).Clinical: 58 artifact cases (1,087 slices); 22 clinical cases with intraoral scans.Simulated:PSNR = 46.1 dB,SSIM = 0.9913Clinical:Likert = 4.46,Dice = 86.3,HD = 11.3Yu et al.(2025) [138]Generating realistic panoramic dental radiographsStable Diffusion 1.5Panoramic radiographsDental X-ray images: 3,657 Segmentation masks: 3,441 annotated imagesDownstream tasks: 2,900 annotated panoramic dental images and 18,240 VQA pairsFID (↓ better): 72.91IS (↑ better): 2.85MS-SSIM (↑ better): 0.63Prediction & Other Clinical Applications (con’t)Yang et al.(2025) [128]Automatically generate high-fidelity, realistic dental crown prosthesis meshes from 3D scans of preparation and antagonist teethDCrownFormer+3D scans3,144 scansCD: 41.004F1-score: 93.1HD: 62.932Yadalam et al.(2025) [141]Predict AIM-2 inflammasome peptide sequences associated with periodontal inflammationTransformerFASTA sequencesUniProt databaseACC: 84SEN: 83AUC: 89.1MCC: 68Zhang et al.(2025) [129]Generate high-resolution, complete, and dense dental point cloudsMPGLNet3D scans16,384 points for complete clouds, 2,048 points for partialCDL1:199.038CDL2: 94.218Ha et al.(2025) [139]Generate T2-weighted (T2w) MR images from proton density-weighted (PDw) images in temporomandibular disorder (TMD) patientsTransUNet-SegGANMRI scans7,226 MR slicesSSIM: 82.28FID: 23.85Ma et al.(2025) [134]Tooth position predictionF-VAMPDental position prediction dataset834 Dental position prediction cases constructed from pathological cases.CSA: 87.67MEt: 0.8338 mmSilva et al.(2024) [142]Evaluate the performance in describing radiolucent lesions and in establishing diagnosesChatGPT-3.5Panoramic radiographs28 casesDescriptive Accuracy Overall Mean Score: 0.56 ± 0.48Differential Diagnosis Accuracy: Rank-1: 25%Rank-2: 57.14%Rank-3: 67.86%Selected Generative Pre-trained Transformer (GPT) Clinical ApplicationsJi et al.(2025) [143]Evaluate the capability of GPT-4 in answering open-ended questions related to oral cancerGPT-4Open-ended questions60 selected from a book titled “100 Questions about Oral Cancer Health KnowledgeEvaluated using a modified Global Quality Scale [1–5]: Most answers were scored at 4 or 5Salmanpour et al.(2025) [144]Evaluate performance in determining the labiolingual localization of maxillary impacted canines and detecting resorption in adjacent incisorsChatGPT-4Panoramic radiographs & CBCT105 patients for localization task; 50 patients for resorption detectionLocalization:ACC: 37.1Resorption Detection:ACC: 46.0Bordin et al.(2025) [145]Evaluate the accuracy and reliability of ChatGPT-3.5 in answering questions related to dental traumaChatGPT 3.5Open-ended questions45 questions generated by ChatGPT-3.5, evaluated by five experienced endodontistsMean scores:Avulsion: 4.40 ± 0.717Intrusion: 4.13 ± 0.794General Dental Trauma: 3.97 ± 0.885Chuang et al.(2024) [146]Evaluate the effectiveness of using GPT-J for prompt generation to create seed data for training RoBERTa-based named entity recognition (NER) modelsGPT-J-6BClinical notes from EDRs8,125 clinical notes focus on periodontitisF1 Score: 0.92–0.97Ameli et al.(2024) [147]Developing a Clinical Decision Support System (CDSS) that classifies the stage of the disease.BERTClinical notes from EDRs309 anonymized patient chartsStaging:ACC: 66Grading:ACC: 72Selected Generative Pre-trained Transformer (GPT) Clinical Applications (con’t)Molena et al.(2024) [148]Evaluate the accuracy, completeness, and reliability of responses generated by ChatGPT 3.5 in dentistryChatGPT 3.5Expert-formulated dental questions30 questionsAccuracy Scores [1–6]: 4.17Completeness Scores [1–3]: 2.07Giannakopoulos et al.(2023) [149]Evaluate and compare the performance of four LLMs in answering dentistry questionsChatGPT-3.5, ChatGPT-4Open-ended, evidence-based clinical dentistry question20 questionsAverage scores [1–10]:ChatGPT-4: 7.2ChatGPT-3.5: 5.9Google Bard: 5.7Microsoft Bing Chat: 5.4ACC, Accuracy; AP: Average Precision; AUC, Area Under the ROC Curve; CBCT, Cone Beam Computed Tomography; CD: Chamfer Distance; CD-L: Chamfer-L Distance; CSA: Correct Spatial Alignment; Dice coefficient; Dice, Sørensen Dice coefficient; DSC, Dice Similarity Coefficient; EMD: Earth Mover’s Distance; FID: Frechet Inception Distance; HD, Hausdorff distance; IoU, Intersection over Union; IS: Inception Score; MAE: mean absolute error; MCC: Matthews correlation coefficient; MEt: Mean; Translation Error; mAP: mean Average Precision; MSE: mean squared error; MS-SSIM: Multi-Scale Structural; Similarity Index Metric; mIoU, mean intersection over union; PA: Pixel Accuracy; PRE, Precision; PSNR: peak signal-to-noise ratio; RMS: Root mean Square; SEN, Sensitivity; SPE, Specificity; SSIM: structural similarity index Characteristics of the clinical generative applications Yang et al. (2024) [118] AP: 13.7 F1-score: 13.6 Fan et al. (2023) [120] MAE: 2.61 (years) AUC: 95.54 Li et al. (2022) [133] ACC: 90.20 AUC: 95.63 SEN: 91.39 SPE: 95.09 F1-score: 90.48 Shen et al. (2023) [122] F-score@1%: 82.88 CD-L2: 0.00122 EMD: 0.0713 Lin et al. (2025) [130] HD90 (XY-axis): 3.50 mm HD90 (XYZ-axis): 4.06 mm Saleh et al. (2024) [123] TM highest accuracy: Tooth 34 (RMS = 0.21) Farook et al. (2023) [140] Yang et al. (2024) [119] AP75: 15.7 F1 Score: 15.6 Hosseinimanesh et al. (2025) [124] Ahmed et al. (2024) [135] Arian et al. (2023) [121] Liu et al. (2023) [127] CBCT Data: 503 scans with 9,651 manually annotated slices IOS Meshes: 28,559 scans Trachoo et al. (2025) [131] Classification accuracy: 86.71(ResNet101V2) Detection average precision: 99.28(RetinaNet) Predicting removal difficulty levels accuracy: 78.99(Vision Transformer) Sivaanpu et al. (2024) [136] MSE: as low as 13.04 PSNR: up to 37.85 dB SSIM: up to 0.98 Lee et al. (2022) [132] Detection: mAP@0.5: 99.0 Extraction Difficulty/IAN injury Prediction: ACC: 83.5/81.1 Alsheghri et al. (2024) [125] CD: 0.137 mm HD: 0.242 mm Liu et al. (2023) [126] CD-L1 (×10− 3): 222 CD-L2 (×10− 3): 194 EMD: 775 P2S: 0.0112 Li et al. (2025) [137] Simulated: 28,036 paired images (45 cases). Clinical: 58 artifact cases (1,087 slices); 22 clinical cases with intraoral scans. Simulated: PSNR = 46.1 dB, SSIM = 0.9913 Clinical: Likert = 4.46, Dice = 86.3, HD = 11.3 Yu et al. (2025) [138] Dental X-ray images: 3,657 Segmentation masks: 3,441 annotated images Downstream tasks: 2,900 annotated panoramic dental images and 18,240 VQA pairs FID (↓ better): 72.91 IS (↑ better): 2.85 MS-SSIM (↑ better): 0.63 Yang et al. (2025) [128] CD: 41.004 F1-score: 93.1 HD: 62.932 Yadalam et al. (2025) [141] ACC: 84 SEN: 83 AUC: 89.1 MCC: 68 Zhang et al. (2025) [129] CDL1:199.038 CDL2: 94.218 Ha et al. (2025) [139] SSIM: 82.28 FID: 23.85 Ma et al. (2025) [134] CSA: 87.67 MEt: 0.8338 mm Silva et al. (2024) [142] Descriptive Accuracy Overall Mean Score: 0.56 ± 0.48 Differential Diagnosis Accuracy: Rank-1: 25% Rank-2: 57.14% Rank-3: 67.86% Ji et al. (2025) [143] Salmanpour et al. (2025) [144] Localization: ACC: 37.1 Resorption Detection: ACC: 46.0 Bordin et al. (2025) [145] Mean scores: Avulsion: 4.40 ± 0.717 Intrusion: 4.13 ± 0.794 General Dental Trauma: 3.97 ± 0.885 Chuang et al. (2024) [146] Ameli et al. (2024) [147] Staging: ACC: 66 Grading: ACC: 72 Molena et al. (2024) [148] Accuracy Scores [1–6]: 4.17 Completeness Scores [1–3]: 2.07 Giannakopoulos et al. (2023) [149] Average scores [1–10]: ChatGPT-4: 7.2 ChatGPT-3.5: 5.9 Google Bard: 5.7 Microsoft Bing Chat: 5.4 ACC, Accuracy; AP: Average Precision; AUC, Area Under the ROC Curve; CBCT, Cone Beam Computed Tomography; CD: Chamfer Distance; CD-L: Chamfer-L Distance; CSA: Correct Spatial Alignment; Dice coefficient; Dice, Sørensen Dice coefficient; DSC, Dice Similarity Coefficient; EMD: Earth Mover’s Distance; FID: Frechet Inception Distance; HD, Hausdorff distance; IoU, Intersection over Union; IS: Inception Score; MAE: mean absolute error; MCC: Matthews correlation coefficient; MEt: Mean; Translation Error; mAP: mean Average Precision; MSE: mean squared error; MS-SSIM: Multi-Scale Structural; Similarity Index Metric; mIoU, mean intersection over union; PA: Pixel Accuracy; PRE, Precision; PSNR: peak signal-to-noise ratio; RMS: Root mean Square; SEN, Sensitivity; SPE, Specificity; SSIM: structural similarity index Transformer-based LLMs in dental clinical applications In addition to domain-specific applications, a subset of studies explored the role of GPT-based language models, which are built on the Transformer architecture, in dental clinical contexts. The remarkable proficiency of LLMs in natural language processing has generated considerable interest in their potential applications across various fields, including clinical dentistry [150]. 7 studies were selected as examples of investigating GPT-3.5, GPT-4, and related models in various dental use cases (Table 3), split into two key categories:Clinical NLP tasks: Domain-adapted Transformer models showed strong performance for structured data extraction. Chuang et al. developed a GPT-J-facilitated RoBERTa model for periodontal diagnosis extraction from electronic dental records (F1 = 0.92–0.97) [146]. Ameli et al.’s BERT-LSTM model outperformed rule-based systems for periodontitis staging and grading [147].General-purpose LLM utility: Performance of commercial models (GPT-3.5/GPT-4) was heterogeneous. Studies reported acceptable utility for dental education and dental trauma Q&A [143, 145], with GPT-4 outperforming other LLMs in evidence-based dental queries [149]. However, critical reliability limitations were consistently noted, including inconsistent radiographic lesion diagnosis [142], poor performance in the impacted canine assessment [144], and unreliable academic reference citation [148]. Clinical NLP tasks: Domain-adapted Transformer models showed strong performance for structured data extraction. Chuang et al. developed a GPT-J-facilitated RoBERTa model for periodontal diagnosis extraction from electronic dental records (F1 = 0.92–0.97) [146]. Ameli et al.’s BERT-LSTM model outperformed rule-based systems for periodontitis staging and grading [147]. Clinical NLP tasks: Domain-adapted Transformer models showed strong performance for structured data extraction. Chuang et al. developed a GPT-J-facilitated RoBERTa model for periodontal diagnosis extraction from electronic dental records (F1 = 0.92–0.97) [146]. Ameli et al.’s BERT-LSTM model outperformed rule-based systems for periodontitis staging and grading [147]. General-purpose LLM utility: Performance of commercial models (GPT-3.5/GPT-4) was heterogeneous. Studies reported acceptable utility for dental education and dental trauma Q&A [143, 145], with GPT-4 outperforming other LLMs in evidence-based dental queries [149]. However, critical reliability limitations were consistently noted, including inconsistent radiographic lesion diagnosis [142], poor performance in the impacted canine assessment [144], and unreliable academic reference citation [148]. General-purpose LLM utility: Performance of commercial models (GPT-3.5/GPT-4) was heterogeneous. Studies reported acceptable utility for dental education and dental trauma Q&A [143, 145], with GPT-4 outperforming other LLMs in evidence-based dental queries [149]. However, critical reliability limitations were consistently noted, including inconsistent radiographic lesion diagnosis [142], poor performance in the impacted canine assessment [144], and unreliable academic reference citation [148]. Quantitative synthesis of model performance This section presents a descriptive quantitative synthesis of the 112 included studies, stratified by core clinical tasks, architecture types, and validation strategies. Formal meta-analysis was not performed a priori due to significant heterogeneity in imaging modalities, target structures, outcome metrics, and study designs, consistent with our PROSPERO-registered protocol. Overall performance by clinical task We summarized the most widely reported performance metrics for the three primary application domains (detailed in Table 4):Segmentation tasks (n = 26) use DSC as the primary endpoint, with performance varying by target structure (Table 4).Detection, classification & localization tasks (n = 54): Core endpoints were ACC and AUC. The overall median ACC was 95.60% (range 72.60%–99.99%), with a median AUC of 0.97 (range 0.87–0.99).Generative & predictive applications (n = 32): Metrics varied by clinical endpoint. For 3D restorative reconstruction, the median Chamfer Distance was 0.137 mm (meeting clinical precision requirements); for dental age estimation, the median mean absolute error (MAE) was 2.61 years.Table 4Summary of Transformer model performance by core clinical taskTask CategorySub-taskNumber of StudiesCore MetricMetric RangeMedianBest-performing Architecture (Reference No.)Segmentation (n = 26)2D/3D Tooth Segmentation16DSC (%)83.86–97.6294.54Trans-VNet [44], SemiTNet [47] AlignNet [57]Mandibular Canal/Bone/Airway Segmentation10DSC (%)79.06–96.8384.43D UNesT [67], Swin UNETR [63]Detection, Classification & Localization (n = 54)Dental Caries Detection/Classification17ACC (%)85.57–98.5396.59DentC_BSFLNet [78], SegFormer [30]Oral Cancer/Precancerous Lesion Classification5ACC (%)88.70–98.5898.58Swin Transformer [96]Multi-disease/Anomaly Classification15ACC (%)92.00–99.0395.6OralTransNet [87], Swin Transformer Ensemble [82]Jaw Lesion/Periapical Disease Detection8AUC0.87–0.990.97CrossViT [92], Swin UNETR [90]Other Anomaly Detection (Impacted teeth, plaque, fluorosis)9F1-score (%)75.79–98.3091.72YOLOv8+RT-DETR [100], CenterFormer [103]Generative & Predictive Applications (n = 32)Dental Age/Sex Estimation2MAE (years) / ACC (%)2.61 years / 95.5%2.61 / 95.5DASE-Net [120]3D Dental Crown/Restoration Generation7Chamfer Distance (mm)0.041–0.1370.071DCrownFormer+ [128], AdaPoinTr [125]Implant Position Prediction2AP13.7–15.714.7ImplantFormer [118]Image Denoising/Artifact Reduction4PSNR (dB) / SSIM37.85 / 0.9837.85 / 0.98HCTSpeckle [136], Semi-SFTrans [137]Dental Large Language Model Applications7ACC / F1-score37.1–96% / 0.92–0.9770.5% / 0.94Fine-tuned RoBERTa [146], GPT-4 [143, 149]DSC: Dice Similarity Coefficient; ACC: Accuracy; AUC: Area Under the Curve; MAE: Mean Absolute Error; AP: Average Precision; PSNR: Peak Signal-to-Noise Ratio; SSIM: Structural Similarity Index Segmentation tasks (n = 26) use DSC as the primary endpoint, with performance varying by target structure (Table 4). Segmentation tasks (n = 26) use DSC as the primary endpoint, with performance varying by target structure (Table 4). Detection, classification & localization tasks (n = 54): Core endpoints were ACC and AUC. The overall median ACC was 95.60% (range 72.60%–99.99%), with a median AUC of 0.97 (range 0.87–0.99). Detection, classification & localization tasks (n = 54): Core endpoints were ACC and AUC. The overall median ACC was 95.60% (range 72.60%–99.99%), with a median AUC of 0.97 (range 0.87–0.99). Generative & predictive applications (n = 32): Metrics varied by clinical endpoint. For 3D restorative reconstruction, the median Chamfer Distance was 0.137 mm (meeting clinical precision requirements); for dental age estimation, the median mean absolute error (MAE) was 2.61 years. Generative & predictive applications (n = 32): Metrics varied by clinical endpoint. For 3D restorative reconstruction, the median Chamfer Distance was 0.137 mm (meeting clinical precision requirements); for dental age estimation, the median mean absolute error (MAE) was 2.61 years. Table 4Summary of Transformer model performance by core clinical taskTask CategorySub-taskNumber of StudiesCore MetricMetric RangeMedianBest-performing Architecture (Reference No.)Segmentation (n = 26)2D/3D Tooth Segmentation16DSC (%)83.86–97.6294.54Trans-VNet [44], SemiTNet [47] AlignNet [57]Mandibular Canal/Bone/Airway Segmentation10DSC (%)79.06–96.8384.43D UNesT [67], Swin UNETR [63]Detection, Classification & Localization (n = 54)Dental Caries Detection/Classification17ACC (%)85.57–98.5396.59DentC_BSFLNet [78], SegFormer [30]Oral Cancer/Precancerous Lesion Classification5ACC (%)88.70–98.5898.58Swin Transformer [96]Multi-disease/Anomaly Classification15ACC (%)92.00–99.0395.6OralTransNet [87], Swin Transformer Ensemble [82]Jaw Lesion/Periapical Disease Detection8AUC0.87–0.990.97CrossViT [92], Swin UNETR [90]Other Anomaly Detection (Impacted teeth, plaque, fluorosis)9F1-score (%)75.79–98.3091.72YOLOv8+RT-DETR [100], CenterFormer [103]Generative & Predictive Applications (n = 32)Dental Age/Sex Estimation2MAE (years) / ACC (%)2.61 years / 95.5%2.61 / 95.5DASE-Net [120]3D Dental Crown/Restoration Generation7Chamfer Distance (mm)0.041–0.1370.071DCrownFormer+ [128], AdaPoinTr [125]Implant Position Prediction2AP13.7–15.714.7ImplantFormer [118]Image Denoising/Artifact Reduction4PSNR (dB) / SSIM37.85 / 0.9837.85 / 0.98HCTSpeckle [136], Semi-SFTrans [137]Dental Large Language Model Applications7ACC / F1-score37.1–96% / 0.92–0.9770.5% / 0.94Fine-tuned RoBERTa [146], GPT-4 [143, 149]DSC: Dice Similarity Coefficient; ACC: Accuracy; AUC: Area Under the Curve; MAE: Mean Absolute Error; AP: Average Precision; PSNR: Peak Signal-to-Noise Ratio; SSIM: Structural Similarity Index Summary of Transformer model performance by core clinical task DSC: Dice Similarity Coefficient; ACC: Accuracy; AUC: Area Under the Curve; MAE: Mean Absolute Error; AP: Average Precision; PSNR: Peak Signal-to-Noise Ratio; SSIM: Structural Similarity Index Subgroup comparison: hybrid CNN–Transformer vs. pure Transformer architectures Head-to-head comparisons in included studies showed consistent advantages of hybrid CNN–Transformer models:Segmentation tasks: Hybrid models achieved a 3.42% higher median DSC than pure Transformer architectures, with the greatest gains in CBCT and panoramic radiograph segmentation, by preserving fine-grained local features while capturing global context.Classification tasks: Hybrid models had a 2.44% higher median ACC than pure Transformer models, with lightweight hybrid designs balancing computational efficiency and diagnostic accuracy.Pure Transformer architectures showed unique strengths in 3D point cloud segmentation and long-sequence generative tasks. Segmentation tasks: Hybrid models achieved a 3.42% higher median DSC than pure Transformer architectures, with the greatest gains in CBCT and panoramic radiograph segmentation, by preserving fine-grained local features while capturing global context. Segmentation tasks: Hybrid models achieved a 3.42% higher median DSC than pure Transformer architectures, with the greatest gains in CBCT and panoramic radiograph segmentation, by preserving fine-grained local features while capturing global context. Classification tasks: Hybrid models had a 2.44% higher median ACC than pure Transformer models, with lightweight hybrid designs balancing computational efficiency and diagnostic accuracy. Classification tasks: Hybrid models had a 2.44% higher median ACC than pure Transformer models, with lightweight hybrid designs balancing computational efficiency and diagnostic accuracy. Pure Transformer architectures showed unique strengths in 3D point cloud segmentation and long-sequence generative tasks. Pure Transformer architectures showed unique strengths in 3D point cloud segmentation and long-sequence generative tasks. Systematic analysis of dataset and validation characteristics This section systematically evaluates dataset attributes, study design, and validation strategies across all 112 included studies, to address gaps in generalizability and clinical translatability (detailed in Table 5).Table 5Summary of dataset and validation characteristics across included studiesCharacteristicSubgroupNumber of Studies (Total n = 112)ProportionDescriptive StatisticsSample SizeAll included studies112100%Median: 1200 cases; IQR: 425–6400 cases2D imaging studies8777.70%Median: 1500 cases; IQR: 506–6400 cases3D imaging/point cloud studies2522.30%Median: 916 scans; IQR: 340–4000 scansSmall dataset (<200 cases)2421.40%Median: 137 cases; IQR: 100–180 casesDataset ProvenancePrivate single-center dataset9786.60%Public/open-access dataset1513.40%Single-center data collection10089.30%Multi-center data collection1210.70%Study DesignRetrospective data collection10896.40%Prospective data collection43.60%Validation StrategyInternal hold-out testing only7163.40%Internal k-fold cross-validation4136.60%Any external validation (independent cohort)3430.40%Multi-center external validation87.10%IQR: Interquartile Range Table 5Summary of dataset and validation characteristics across included studiesCharacteristicSubgroupNumber of Studies (Total n = 112)ProportionDescriptive StatisticsSample SizeAll included studies112100%Median: 1200 cases; IQR: 425–6400 cases2D imaging studies8777.70%Median: 1500 cases; IQR: 506–6400 cases3D imaging/point cloud studies2522.30%Median: 916 scans; IQR: 340–4000 scansSmall dataset (<200 cases)2421.40%Median: 137 cases; IQR: 100–180 casesDataset ProvenancePrivate single-center dataset9786.60%Public/open-access dataset1513.40%Single-center data collection10089.30%Multi-center data collection1210.70%Study DesignRetrospective data collection10896.40%Prospective data collection43.60%Validation StrategyInternal hold-out testing only7163.40%Internal k-fold cross-validation4136.60%Any external validation (independent cohort)3430.40%Multi-center external validation87.10%IQR: Interquartile Range Summary of dataset and validation characteristics across included studies IQR: Interquartile Range 3.4.1 Dataset size and distribution Sample sizes varied widely across included studies:Overall median sample size: 1200 imaging cases/scans (Interquartile Range, Interquartile Range (IQR): 425–6400 cases; range: 100– 56,340 images).2D imaging studies (panoramic radiographs, intraoral photos, etc., n = 87): median sample size 1500 cases (IQR: 506–6400 cases).3D imaging studies (CBCT, intraoral 3D scans, etc., n = 25): median sample size 916 scans/models (IQR: 340–4000 scans).24 of 112 studies used small datasets with <200 cases, mostly focused on rare dental anomalies or pathologies. Overall median sample size: 1200 imaging cases/scans (Interquartile Range, Interquartile Range (IQR): 425–6400 cases; range: 100– 56,340 images). Overall median sample size: 1200 imaging cases/scans (Interquartile Range, Interquartile Range (IQR): 425–6400 cases; range: 100– 56,340 images). 2D imaging studies (panoramic radiographs, intraoral photos, etc., n = 87): median sample size 1500 cases (IQR: 506–6400 cases). 2D imaging studies (panoramic radiographs, intraoral photos, etc., n = 87): median sample size 1500 cases (IQR: 506–6400 cases). 3D imaging studies (CBCT, intraoral 3D scans, etc., n = 25): median sample size 916 scans/models (IQR: 340–4000 scans). 3D imaging studies (CBCT, intraoral 3D scans, etc., n = 25): median sample size 916 scans/models (IQR: 340–4000 scans). 24 of 112 studies used small datasets with <200 cases, mostly focused on rare dental anomalies or pathologies. 24 of 112 studies used small datasets with <200 cases, mostly focused on rare dental anomalies or pathologies. 3.4.2 Dataset source and multicenter status Public vs. private datasets: Only 15 of 112 studies used fully or partially public open-access dental imaging datasets. The vast majority (86.6%, 97 of 112) relied on private, institutional single-center datasets with no public access. No included studies provided open access to both the full dataset and model code.Single-center vs. multi-center data: 100 of 112 studies used data collected from a single clinical center. Only 12 studies used multi-center datasets, with just 4 studies incorporating data from ≥3 clinical sites. Public vs. private datasets: Only 15 of 112 studies used fully or partially public open-access dental imaging datasets. The vast majority (86.6%, 97 of 112) relied on private, institutional single-center datasets with no public access. No included studies provided open access to both the full dataset and model code. Public vs. private datasets: Only 15 of 112 studies used fully or partially public open-access dental imaging datasets. The vast majority (86.6%, 97 of 112) relied on private, institutional single-center datasets with no public access. No included studies provided open access to both the full dataset and model code. Single-center vs. multi-center data: 100 of 112 studies used data collected from a single clinical center. Only 12 studies used multi-center datasets, with just 4 studies incorporating data from ≥3 clinical sites. Single-center vs. multi-center data: 100 of 112 studies used data collected from a single clinical center. Only 12 studies used multi-center datasets, with just 4 studies incorporating data from ≥3 clinical sites. Study design and validation strategy Retrospective vs. prospective design: 108 of 112 studies used retrospectively collected clinical imaging data. Only 4 studies used prospectively collected data, with no randomized controlled trials (RCTs) evaluating real-world clinical impact.Validation methodology: 71 of 112 studies used a single train-validation-test hold-out split within a single cohort; 41 used k-fold cross-validation. Critically, only 34 of the studies reported any external validation using an independent, unseen cohort, and only 8 studies performed multi-center external validation. Retrospective vs. prospective design: 108 of 112 studies used retrospectively collected clinical imaging data. Only 4 studies used prospectively collected data, with no randomized controlled trials (RCTs) evaluating real-world clinical impact. Retrospective vs. prospective design: 108 of 112 studies used retrospectively collected clinical imaging data. Only 4 studies used prospectively collected data, with no randomized controlled trials (RCTs) evaluating real-world clinical impact. Validation methodology: 71 of 112 studies used a single train-validation-test hold-out split within a single cohort; 41 used k-fold cross-validation. Critically, only 34 of the studies reported any external validation using an independent, unseen cohort, and only 8 studies performed multi-center external validation. Validation methodology: 71 of 112 studies used a single train-validation-test hold-out split within a single cohort; 41 used k-fold cross-validation. Critically, only 34 of the studies reported any external validation using an independent, unseen cohort, and only 8 studies performed multi-center external validation. Discussion Main findings and interpretation This study evaluated the last five years of literature research on the applications of dental fields using Transformer-based AI models, using a systematic review. According to the results of this evaluation, some findings that need to be discussed have emerged. Even if it is not as many as the applications in the field of medicine, the results of the searches made in five databases show the records of 4862 for the last five years, which shows that Transformer is experiencing a golden age in the applications in dentistry. Over the years, Transformer has grown in popularity as a research topic in dental applications. We also found a significant part of the research identified is not from dentistry but from technical sciences and has been published differently. In dental image analysis, segmentation and object detection often serve as essential preprocessing steps prior to classification or diagnostic analysis. Transformer-based segmentation models are primarily applied to CBCT, periapical and panoramic radiographs, and 3D dental models, owing to the substantial data and annotation requirements inherent to these tasks. The dominance of hybrid CNN–Transformer architectures stems from their complementary strengths: CNNs excel at extracting fine-grained local spatial features, while Transformers address CNNs’ core limitation by modeling long-range global context. This synergy, which preserves anatomical detail while enabling holistic image interpretation, is supported by our head-to-head comparison data as the key driver of performance gains over conventional pure CNN models in most dental tasks. Sample size emerged as the strongest predictor of model performance and generalizability across all included studies. The median sample size was 1200 cases (IQR: 425–6400), with 21.4% of studies using small datasets (<200 cases) for rare conditions. We observed consistent performance gaps: segmentation models with >1000 cases achieved 7.3% higher median DSC than small datasets; classification models with >5000 cases showed 6.5% higher accuracy; and 3D generative models with >500 scans had 0.118 mm better median Chamfer Distance. Critically, small datasets had a 12.3% median performance drop in external validation, compared to only 4.7% for large datasets, due to overfitting to dataset-specific noise. The scarcity of large annotated dental datasets stems from high expert annotation costs, privacy restrictions, and limited rare disease cases. Future research should prioritize multi-center data sharing, self-supervised learning, and transfer learning to mitigate these limitations. In the domain of 3D image analysis, the literature remains relatively sparse regarding robust techniques for the representation and processing of 3D data, raising concerns about the optimization of performance in this area. CNNs have long dominated computer vision benchmarks due to their capacity for local spatial feature extraction, translational invariance, and hierarchical data representation, making them well-suited for large-scale image datasets. However, this local focus also limits the ability of CNNs to model long-range dependencies within the data. In contrast, ViTs, a more recent self-supervised architecture, address this shortcoming through self-attention mechanisms that attend to all image patches simultaneously. This approach enables the model to capture global relationships and fine-grained details, which are particularly valuable for complex image interpretation tasks. Studies combining CNNs and Transformers in a unified architecture reflect an ongoing effort to harness their complementary strengths across classification and other clinical applications. Regarding classification tasks and other generative clinical applications, dental caries emerged as the most frequently studied anomaly in the literature (n = 17). Nevertheless, the majority of other dental conditions were addressed in only one or two studies, underscoring the need for broader exploration in underrepresented areas. While the high performance reported in many of the reviewed studies demonstrates the potential reliability of Transformer-based models, several studies lacked the use of clinically appropriate performance metrics and did not include validation through additional clinical tests. Although most research emphasized the role of AI-supported expert systems as decision-support tools for clinicians, many failed to implement evaluation protocols aligned with this intended purpose. Additionally, while commercial tools such as ChatGPT have demonstrated success in natural language tasks, their clinical applicability remains unexplored. Integrating LLMs’ capabilities with Transformer-based imaging models may offer a promising direction for future clinical use cases. Two primary deployment paradigms are manily focused in recent years’ studies: automated dental diagnosis and cross-modal/multimodal dental analysis. When augmented with a cross-modal encoder, a single LLM can integrate and reason over multi-source data (e.g., textual notes, radiographs, and clinical images) to facilitate complex clinical decision-making. Studies included illustrated the potential of fully integrated, multi-modal LLM systems for clinical deployment. A significant limitation of this, however, lies in the current lack of clinical specificity. A critical pathway to overcoming this limitation is the fine-tuning of base models with domain-specific knowledge derived from dental textbooks, curated research literature, and de-identified patient records. This process enhances diagnostic accuracy by embedding domain-specific patterns, terminology, and context, thereby enabling the generation of clinically relevant and insightful outputs. The integration of Transformer-based imaging models and LLM frameworks provides a potential technical path for multi-modal dental decision support systems. However, current research remains in the preclinical exploratory stage, with major barriers to clinical application, including the lack of prospective validation, insufficient data privacy guarantees, inherent model biases, and the absence of standardized clinical evaluation protocols. Future research must therefore focus on these areas to translate the theoretical promise of LLMs into safe, effective, and equitable clinical practice. Comparison of 2D and 3D Transformer approaches in dental CBCT analysis In CBCT analysis, this study mainly focused on 2D slice-based Transformer analysis methods based on a comprehensive consideration of clinical practice, computational efficiency, data availability, and task characteristics. We systematically compared the two approaches across six core dimensions:Clinical practicality: 2D slice-based approaches align perfectly with clinicians’ daily workflow of browsing axial, coronal, and sagittal slices, producing directly interpretable and verifiable results. In contrast, 3D visualization requires specialized software and additional training, resulting in limited adoption in routine clinical practice.Computational efficiency: 2D models have low computational complexity, enabling fast training and real-time inference on standard clinical workstations (median inference time < 1 second per image).Data availability and annotation cost: Large-scale 2D dental datasets are more readily available, and 2D slice annotation is efficient. 3D volumetric pixel-level annotation is extremely labor-intensive, and very few public 3D dental datasets exist.Model maturity and pre-training: Mature 2D Vision Transformer architectures (e.g., Swin Transformer, SegFormer) with extensive pre-trained models are widely available. 3D Transformer technology is less developed, with limited pre-trained models for medical imaging.Task performance: 2D models achieved excellent performance for most diagnostic tasks (segmentation, detection, classification). 3D models only outperformed 2D approaches in complex 3D spatial tasks such as 3D tooth reconstruction and volumetric measurement.Generalizability: 2D models demonstrated better generalizability across different imaging devices and populations. 3D models are more prone to overfitting on small datasets. Clinical practicality: 2D slice-based approaches align perfectly with clinicians’ daily workflow of browsing axial, coronal, and sagittal slices, producing directly interpretable and verifiable results. In contrast, 3D visualization requires specialized software and additional training, resulting in limited adoption in routine clinical practice. Clinical practicality: 2D slice-based approaches align perfectly with clinicians’ daily workflow of browsing axial, coronal, and sagittal slices, producing directly interpretable and verifiable results. In contrast, 3D visualization requires specialized software and additional training, resulting in limited adoption in routine clinical practice. Computational efficiency: 2D models have low computational complexity, enabling fast training and real-time inference on standard clinical workstations (median inference time < 1 second per image). Computational efficiency: 2D models have low computational complexity, enabling fast training and real-time inference on standard clinical workstations (median inference time < 1 second per image). Data availability and annotation cost: Large-scale 2D dental datasets are more readily available, and 2D slice annotation is efficient. 3D volumetric pixel-level annotation is extremely labor-intensive, and very few public 3D dental datasets exist. Data availability and annotation cost: Large-scale 2D dental datasets are more readily available, and 2D slice annotation is efficient. 3D volumetric pixel-level annotation is extremely labor-intensive, and very few public 3D dental datasets exist. Model maturity and pre-training: Mature 2D Vision Transformer architectures (e.g., Swin Transformer, SegFormer) with extensive pre-trained models are widely available. 3D Transformer technology is less developed, with limited pre-trained models for medical imaging. Model maturity and pre-training: Mature 2D Vision Transformer architectures (e.g., Swin Transformer, SegFormer) with extensive pre-trained models are widely available. 3D Transformer technology is less developed, with limited pre-trained models for medical imaging. Task performance: 2D models achieved excellent performance for most diagnostic tasks (segmentation, detection, classification). 3D models only outperformed 2D approaches in complex 3D spatial tasks such as 3D tooth reconstruction and volumetric measurement. Task performance: 2D models achieved excellent performance for most diagnostic tasks (segmentation, detection, classification). 3D models only outperformed 2D approaches in complex 3D spatial tasks such as 3D tooth reconstruction and volumetric measurement. Generalizability: 2D models demonstrated better generalizability across different imaging devices and populations. 3D models are more prone to overfitting on small datasets. Generalizability: 2D models demonstrated better generalizability across different imaging devices and populations. 3D models are more prone to overfitting on small datasets. It should be noted that 3D Transformer approaches have unique advantages for tasks requiring precise 3D spatial information, such as 3D tooth segmentation, implant planning, and craniomaxillofacial reconstruction. Recent advances in lightweight 3D Transformers (e.g., Swin Transformer 3D, UNETR) and self-supervised learning are addressing the computational and data limitations of 3D methods, making them increasingly promising for future clinical applications. Limitations and future recommendations This systematic review has several limitations. The inclusion and exclusion criteria, while clearly defined, restricted the study selection to peer-reviewed journal articles, excluding conference proceedings, preprints, and book chapters. Additionally, we only included English-language publications, which may introduce language bias and potentially exclude relevant studies published in other languages. Some potentially relevant studies were not included due to limited access or incomplete data incompatible with our summary tables. The data extraction was independently performed by two reviewers, with all discrepancies resolved through consensus review by a third senior reviewer with clinical expertise in dentistry, ensuring the accuracy and reliability of the extracted data. As this was a traditional systematic review, no meta-analysis was conducted. Consequently, the findings are presented in a narrative format, organized according to a framework designed to guide future research. Despite these limitations, this review offers a structured overview of current applications of Transformer-based models in dentistry and identifies key directions for advancing their clinical utility. The following points outline key perspectives on the potential applications and future directions of Transformer-based models in dental imaging and diagnostics:Advancements in Transformer Architectures: Hybrid CNN–Transformer designs, the mainstream paradigm validated in this review, leverage the complementary strengths of CNNs and Transformers to boost diagnostic performance. Future research may explore lightweight, task-optimized hybrid architectures for broader clinical translation.Critical Barriers to Clinical Translation: The included studies are predominantly retrospective, single-center, and based on offline annotated datasets, with very limited prospective validation. The included studies are predominantly retrospective, single-center, and based on offline annotated datasets, with very limited prospective validation. Additionally, while 2D slice-based approaches are currently dominant in dental AI research and clinical practice due to their efficiency and clinical compatibility, they cannot fully capture complex 3D anatomical relationships. Future research should focus on developing lightweight 3D Transformer architectures and hybrid 2D-3D fusion models that combine the advantages of both approaches, as well as constructing large-scale public 3D dental imaging datasets to support the development of robust and clinically translatable AI systems. No studies have completed regulatory approval or real-world clinical workflow integration evaluation. Data siloing blocks independent verification, slows scientific progress, and worsens overfitting and inflated performance estimates; models trained on single‑center data learn institution‑specific biases rather than generalizable features, causing a clinically meaningful performance drop when deployed to new populations, scanners, or protocols. To enable safe, robust, and clinically translatable AI, no transformer‑based dental system has yet gained regulatory clearance or real‑world workflow validation. Future work must prioritize: 1) large‑scale prospective multi‑center trials to validate real‑world performance and safety; 2) standardized evaluation protocols aligned with routine clinical practice (including head‑to‑head comparisons with clinicians); 3) full regulatory and ethical review for decision‑support tools; 4) assessment of impacts on clinical efficiency, diagnostic consistency, costs, and patient outcomes; 5) lightweight 3D transformers and hybrid 2D–3D fusion models; 6) large public 3D dental imaging datasets; and 7) open data and open code to support reproducibility and cumulative progress.Challenges in Data Collection and Annotation: A major barrier to the widespread application of Transformer-based models in dentistry is the scarcity of large, well-annotated datasets. Compared to other medical imaging fields, obtaining annotated dental data is particularly challenging. Annotation requires domain-specific expertise and is both labor-intensive and time-consuming, often necessitating the involvement of experienced dental professionals. This challenge is further compounded in rare or complex cases, where reliable annotation may not be feasible. Addressing this limitation will be essential for training robust, generalizable Transformer models in dental applications.Exploration of Multi-modal Fusion Capabilities: One of the most promising yet underexplored features of Transformer architecture is their ability to integrate multi-modal data sources. In dental imaging, this capability opens the door to comprehensive diagnostic frameworks that combine multiple imaging modalities, such as radiographs, CBCT scans, and clinical records, into a unified analytical model. Future research can prioritize the development of such multi-modal systems, which have the potential to improve diagnostic accuracy, enhance decision-making, and reduce reliance on single-modality data. Additionally, integrating NLP capabilities with Transformer-based imaging models may facilitate clinical interpretation and workflow integration. While commercial tools such as ChatGPT have demonstrated effectiveness in language tasks, their direct clinical applicability remains limited. A fusion of NLP and imaging within Transformer frameworks represents a promising avenue for developing holistic, AI-driven diagnostic support systems in dentistry.K-fold cross-validation: A major methodological limitation across the included literature is the over-reliance on single train-validation-test hold-out splits. While this approach is computationally efficient and widely used in preliminary AI research, it is highly susceptible to data partitioning bias and may significantly overestimate model performance. In contrast, k-fold cross-validation, which was used in only one-third of the included studies, provides a more robust and unbiased estimate of model generalization by averaging performance across multiple independent splits. Future dental AI research should prioritize k-fold cross-validation as a standard methodological practice, particularly for studies reporting preliminary performance on single-center datasets. For studies with sufficient sample size, nested cross-validation is further recommended to avoid overfitting during hyperparameter tuning.Model Explainability and Clinical Trust: A key limitation of dental Transformer research is the near absence of model explainability analysis. Most studies fail to link model attention or feature importance to clinical/pathological features, nor evaluate the clinical relevance of explainable outputs or clinicians’ practical use of such information. The “black-box” nature blocks clinical adoption, as dentists require transparent reasoning to verify decisions, detect errors, and avoid diagnostic biases that may cause care disparities. Meanwhile, regulators (FDA, CE) now mandate explainability for AI medical devices. Notably, Transformers’ inherent self-attention enables direct and reliable visualization of focused image regions, an explainable advantage over CNNs that remains untapped in dentistry. Future work must embed explainability as a core design requirement using Grad-CAM, attention maps, SHAP/LIME, concept-based interpretation, and counterfactual explanations. These methods will boost clinical trust, regulatory compliance, diagnostic accuracy, and clinical translation. Advancements in Transformer Architectures: Hybrid CNN–Transformer designs, the mainstream paradigm validated in this review, leverage the complementary strengths of CNNs and Transformers to boost diagnostic performance. Future research may explore lightweight, task-optimized hybrid architectures for broader clinical translation. Advancements in Transformer Architectures: Hybrid CNN–Transformer designs, the mainstream paradigm validated in this review, leverage the complementary strengths of CNNs and Transformers to boost diagnostic performance. Future research may explore lightweight, task-optimized hybrid architectures for broader clinical translation. Critical Barriers to Clinical Translation: The included studies are predominantly retrospective, single-center, and based on offline annotated datasets, with very limited prospective validation. The included studies are predominantly retrospective, single-center, and based on offline annotated datasets, with very limited prospective validation. Additionally, while 2D slice-based approaches are currently dominant in dental AI research and clinical practice due to their efficiency and clinical compatibility, they cannot fully capture complex 3D anatomical relationships. Future research should focus on developing lightweight 3D Transformer architectures and hybrid 2D-3D fusion models that combine the advantages of both approaches, as well as constructing large-scale public 3D dental imaging datasets to support the development of robust and clinically translatable AI systems. No studies have completed regulatory approval or real-world clinical workflow integration evaluation. Data siloing blocks independent verification, slows scientific progress, and worsens overfitting and inflated performance estimates; models trained on single‑center data learn institution‑specific biases rather than generalizable features, causing a clinically meaningful performance drop when deployed to new populations, scanners, or protocols. To enable safe, robust, and clinically translatable AI, no transformer‑based dental system has yet gained regulatory clearance or real‑world workflow validation. Future work must prioritize: 1) large‑scale prospective multi‑center trials to validate real‑world performance and safety; 2) standardized evaluation protocols aligned with routine clinical practice (including head‑to‑head comparisons with clinicians); 3) full regulatory and ethical review for decision‑support tools; 4) assessment of impacts on clinical efficiency, diagnostic consistency, costs, and patient outcomes; 5) lightweight 3D transformers and hybrid 2D–3D fusion models; 6) large public 3D dental imaging datasets; and 7) open data and open code to support reproducibility and cumulative progress. Critical Barriers to Clinical Translation: The included studies are predominantly retrospective, single-center, and based on offline annotated datasets, with very limited prospective validation. The included studies are predominantly retrospective, single-center, and based on offline annotated datasets, with very limited prospective validation. Additionally, while 2D slice-based approaches are currently dominant in dental AI research and clinical practice due to their efficiency and clinical compatibility, they cannot fully capture complex 3D anatomical relationships. Future research should focus on developing lightweight 3D Transformer architectures and hybrid 2D-3D fusion models that combine the advantages of both approaches, as well as constructing large-scale public 3D dental imaging datasets to support the development of robust and clinically translatable AI systems. No studies have completed regulatory approval or real-world clinical workflow integration evaluation. Data siloing blocks independent verification, slows scientific progress, and worsens overfitting and inflated performance estimates; models trained on single‑center data learn institution‑specific biases rather than generalizable features, causing a clinically meaningful performance drop when deployed to new populations, scanners, or protocols. To enable safe, robust, and clinically translatable AI, no transformer‑based dental system has yet gained regulatory clearance or real‑world workflow validation. Future work must prioritize: 1) large‑scale prospective multi‑center trials to validate real‑world performance and safety; 2) standardized evaluation protocols aligned with routine clinical practice (including head‑to‑head comparisons with clinicians); 3) full regulatory and ethical review for decision‑support tools; 4) assessment of impacts on clinical efficiency, diagnostic consistency, costs, and patient outcomes; 5) lightweight 3D transformers and hybrid 2D–3D fusion models; 6) large public 3D dental imaging datasets; and 7) open data and open code to support reproducibility and cumulative progress. Challenges in Data Collection and Annotation: A major barrier to the widespread application of Transformer-based models in dentistry is the scarcity of large, well-annotated datasets. Compared to other medical imaging fields, obtaining annotated dental data is particularly challenging. Annotation requires domain-specific expertise and is both labor-intensive and time-consuming, often necessitating the involvement of experienced dental professionals. This challenge is further compounded in rare or complex cases, where reliable annotation may not be feasible. Addressing this limitation will be essential for training robust, generalizable Transformer models in dental applications. Challenges in Data Collection and Annotation: A major barrier to the widespread application of Transformer-based models in dentistry is the scarcity of large, well-annotated datasets. Compared to other medical imaging fields, obtaining annotated dental data is particularly challenging. Annotation requires domain-specific expertise and is both labor-intensive and time-consuming, often necessitating the involvement of experienced dental professionals. This challenge is further compounded in rare or complex cases, where reliable annotation may not be feasible. Addressing this limitation will be essential for training robust, generalizable Transformer models in dental applications. Exploration of Multi-modal Fusion Capabilities: One of the most promising yet underexplored features of Transformer architecture is their ability to integrate multi-modal data sources. In dental imaging, this capability opens the door to comprehensive diagnostic frameworks that combine multiple imaging modalities, such as radiographs, CBCT scans, and clinical records, into a unified analytical model. Future research can prioritize the development of such multi-modal systems, which have the potential to improve diagnostic accuracy, enhance decision-making, and reduce reliance on single-modality data. Additionally, integrating NLP capabilities with Transformer-based imaging models may facilitate clinical interpretation and workflow integration. While commercial tools such as ChatGPT have demonstrated effectiveness in language tasks, their direct clinical applicability remains limited. A fusion of NLP and imaging within Transformer frameworks represents a promising avenue for developing holistic, AI-driven diagnostic support systems in dentistry. Exploration of Multi-modal Fusion Capabilities: One of the most promising yet underexplored features of Transformer architecture is their ability to integrate multi-modal data sources. In dental imaging, this capability opens the door to comprehensive diagnostic frameworks that combine multiple imaging modalities, such as radiographs, CBCT scans, and clinical records, into a unified analytical model. Future research can prioritize the development of such multi-modal systems, which have the potential to improve diagnostic accuracy, enhance decision-making, and reduce reliance on single-modality data. Additionally, integrating NLP capabilities with Transformer-based imaging models may facilitate clinical interpretation and workflow integration. While commercial tools such as ChatGPT have demonstrated effectiveness in language tasks, their direct clinical applicability remains limited. A fusion of NLP and imaging within Transformer frameworks represents a promising avenue for developing holistic, AI-driven diagnostic support systems in dentistry. K-fold cross-validation: A major methodological limitation across the included literature is the over-reliance on single train-validation-test hold-out splits. While this approach is computationally efficient and widely used in preliminary AI research, it is highly susceptible to data partitioning bias and may significantly overestimate model performance. In contrast, k-fold cross-validation, which was used in only one-third of the included studies, provides a more robust and unbiased estimate of model generalization by averaging performance across multiple independent splits. Future dental AI research should prioritize k-fold cross-validation as a standard methodological practice, particularly for studies reporting preliminary performance on single-center datasets. For studies with sufficient sample size, nested cross-validation is further recommended to avoid overfitting during hyperparameter tuning. K-fold cross-validation: A major methodological limitation across the included literature is the over-reliance on single train-validation-test hold-out splits. While this approach is computationally efficient and widely used in preliminary AI research, it is highly susceptible to data partitioning bias and may significantly overestimate model performance. In contrast, k-fold cross-validation, which was used in only one-third of the included studies, provides a more robust and unbiased estimate of model generalization by averaging performance across multiple independent splits. Future dental AI research should prioritize k-fold cross-validation as a standard methodological practice, particularly for studies reporting preliminary performance on single-center datasets. For studies with sufficient sample size, nested cross-validation is further recommended to avoid overfitting during hyperparameter tuning. Model Explainability and Clinical Trust: A key limitation of dental Transformer research is the near absence of model explainability analysis. Most studies fail to link model attention or feature importance to clinical/pathological features, nor evaluate the clinical relevance of explainable outputs or clinicians’ practical use of such information. The “black-box” nature blocks clinical adoption, as dentists require transparent reasoning to verify decisions, detect errors, and avoid diagnostic biases that may cause care disparities. Meanwhile, regulators (FDA, CE) now mandate explainability for AI medical devices. Notably, Transformers’ inherent self-attention enables direct and reliable visualization of focused image regions, an explainable advantage over CNNs that remains untapped in dentistry. Future work must embed explainability as a core design requirement using Grad-CAM, attention maps, SHAP/LIME, concept-based interpretation, and counterfactual explanations. These methods will boost clinical trust, regulatory compliance, diagnostic accuracy, and clinical translation. Model Explainability and Clinical Trust: A key limitation of dental Transformer research is the near absence of model explainability analysis. Most studies fail to link model attention or feature importance to clinical/pathological features, nor evaluate the clinical relevance of explainable outputs or clinicians’ practical use of such information. The “black-box” nature blocks clinical adoption, as dentists require transparent reasoning to verify decisions, detect errors, and avoid diagnostic biases that may cause care disparities. Meanwhile, regulators (FDA, CE) now mandate explainability for AI medical devices. Notably, Transformers’ inherent self-attention enables direct and reliable visualization of focused image regions, an explainable advantage over CNNs that remains untapped in dentistry. Future work must embed explainability as a core design requirement using Grad-CAM, attention maps, SHAP/LIME, concept-based interpretation, and counterfactual explanations. These methods will boost clinical trust, regulatory compliance, diagnostic accuracy, and clinical translation. Conclusions This systematic review highlights the emerging role of Transformer-based models in the field of dental imaging and diagnostics, reflecting a broader shift toward advanced deep learning techniques in oral healthcare. Transformers have demonstrated considerable potential in tasks involving dental structures such as teeth and bones, outperforming traditional convolutional architectures in capturing global contextual information. These technologies have the potential not only to alleviate the clinical workload but also to democratize access to high-quality diagnostic services by reducing dependence on specialist expertise and associated costs. While deep learning has already shown promising results in the diagnosis of dental anomalies and diseases, the integration of Transformer models in this domain is still in its early stages. To realize these benefits, future research should prioritize the development of clinically validated Transformer-based systems. This includes establishing standardized datasets, benchmarking protocols, robust performance metrics, and ensuring model interpretability and generalizability across diverse populations and imaging modalities. As the field continues to mature, the continued expansion and refinement of Transformer applications hold great promise for advancing diagnostic precision, treatment planning, and ultimately, patient outcomes in oral and dental healthcare. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1