Open research questions in Dental Radiography and Imaging

What the literature leaves open

• Future research will focus on addressing current limitations, such as data privacy concerns and the need for clinical validation, ensuring that AI tools are not only effective but also safe and reliable [22-24].

  Artificial intelligence in endodontics: A comprehensive review of applications and advancement · 2026 · DOI

• raises important ethical and medico-legal concerns. In situations where AI-generated outputs contribute to an incorrect regarding professional retreatment decision, questions accountability, standard-of-care compliance become increasingly relevant. This dimension further reinforces that such systems must remain under strict clinician supervision [8]. informed consent, and Beyond direct clinical implications, the uncritical use of LLMs may also shape diagnostic reasoning patterns among undergraduate students, residents, and early-career clinicians. In Endodontic education, excessive reliance on AI-generated recommendations without adequate critical appraisal may inadvertently weaken the development of independent diagnostic reasoning, radiographic interpretation skills, and biologically grounded clinical judgment [4, 5, 8]. Although AI and LLMs have the potential to enhance endodontic diagnosis, treatment planning, and educational support, their integration into clinical practice should remain strictly adjunctive and continuously guided by expert professional [1, 5, 6]. Until robust validation frameworks, factual verification protocols, and clinically tested multimodal models become widely available, the indiscriminate adoption of these technologies may compromise diagnostic accuracy, weaken clinical accountability, and ultimately expose patients to inappropriate therapeutic decisions [8]. Therefore, judgment the future of AI in Endodontics should be defined not solely by technological sophistication but by its ability to provide transparent, reliable, and biologically sound support that enhances diagnostic safety and strengthens responsible clinical decision-making [5, 7, 9].

  Artificial intelligence in endodontic decision-making: hallucinations and emerging challenges for clinical practice · 2026 · DOI

• Future developments in dental Artificial Intelligence are expected to focus increasingly on transparency, fairness, personalization, and human-centered clinical support systems. improves clinician decision-making, 2.

  Artificial Intelligence in dentistry: The black box problem, ethical challenges, and the role of explainable Artificial Intelligence—A narrative review · 2026 · DOI

• The evolution of AI in prosthodontics can be ca- tegorized into immediate clinical refinements and long-term transformative shifts. In the near term (next 5 years), research priorities are anticipated to focus on the clinical validation of automated design tools and the integration of AI into chairside workflows to enhance diagnostic precision. These refinements include the development of intuitive CAD interfaces that require minimal manual inter- vention. In the long term (5–10 years), the focus will likely shift toward robotics and autonomous systems. While currently limited by high costs and technical complexity, robot-assisted tooth preparati- on represents a significant frontier. These systems promise to deliver a level of precision and repeata- bility that exceeds human capability, particularly in complex full-mouth rehabilitations. However, these technological advances must be guided by a robust ethical governance framework. As AI and robotics take on other autonomous roles, the “human-in-the- loop” principle remains non-negotiable. Clinician oversight is essential to ensure that AI-generated de- cisions align with the patient’s unique biological ne- eds and personal values. Future research must prio- ritize not only technical accuracy but also the 125 development of ethical guidelines regarding data privacy, algorithmic transparency, and the legal accountability of AI-driven clinical outcomes. Such guidelines must ensure that AI remains a transpa- rent and supportive adjunct rather than an opaque decision-maker, which will be the defining challen- ge of the next decade in prosthodontics. CONCLUSION This review demonstrates that while AI is a powerful adjunct in prosthodontics, its clinical ma- turity remains domain specific. AI-assisted diagnos- tic tools and margin detection systems have shown promising levels of precision in retrospective and experimental approaches, suggesting they are suita- ble for integration into digital workflows to enhance standardization and efficiency. However, generati- ve applications like AI-driven smile design and GAN- based prosthesis fabrication remain in the experi- mental phase, requiring further clinical validation and large-scale, standardized datasets before they can be considered routine clinical tools. The transi- tion toward AI-integrated prosthodontics must be guided by the “human-in-the-loop” principle: AI handles quantifiable, data-heavy tasks, while the clinician retains the final authority on qualitative and ethical decisions. Current evidence suggests that AI can be as a de- cision-support system that may reduce procedural time and human error while remaining vigilant about the current limitations in algorithmic trans- parency and the necessity of clinical oversight. Fu- ture progress will depend on a collaborative synergy between technological innovation and rigorous cli- nical validation to ensure that AI serves as a reliable cornerstone of personalized prosthodontic care.

  Application of artificial intelligence in prosthodontics: a narrative review · 2026 · DOI

• The narrative methodology is inherently subject to selection bias, and no formal quality assessment of the included studies was performed. The included studies are heterogeneous in terms of dosage, route of administration, timing of administration, and outcome measurement. This heterogeneity limits direct comparison between studies. Future systematic reviews and meta-analyses using standardized protocols are needed to strengthen the evidence base.

  ERIOPERATIVE CORTICOSTEROID USE IN MANDIBULAR THIRD MOLAR SURGERY: A NARRATIVE REVIEW OF EFFECTS ON POSTOPERATIVE PAIN, EDEMA, AND TRISMUS · 2026 · DOI

• , microdontia/macrodontia) that fall outside the primary distribution of our training set—has not been evaluated and represents a known boundary of the current approach. Its efficacy and accuracy in pedi- atric populations with mixed or primary (deciduous) dentition have not been established. However, its perfor- mance on non-standard images—such as those taken from oblique angles, with varying camera-to-subject distances, under partial occlusion, or using different camera systems—has not been established. Third, regarding clinical precision, although the model demonstrated satisfactory overall estimation perfor- mance, its accuracy in posterior tooth regions remains insufficient to meet high-precision requirements for applications such as prosthetic rehabilitation and dental implantation.

  Automatic estimation of single-tooth width from standardized two-dimensional occlusal photographs using deep learning · 2026 · DOI

• This practice is consistent with the study by Lawani et al,10 indicating a gap in knowledge of the current standards in radiation protection, as lead aprons and thyroid collars are reported to introduce artefacts into dental images by obstructing the primary beam resulting in repeat examinations and do not protect against internal scatter radiation.

  Radiation protection in dental radiology: Compliance, continuing education and equipment audit in imaging clinics · 2026 · DOI

• Yoon et al. [32] Akadiri et al. [33] Pais et al. [34] AUROC (model 1 training): 0.9191, AUROC (model 1 cross-validation): 0.8289; AUROC (model 2 training): 0.9263, AUROC (model 2 cross-validation): 0.8415 (H2O AutoML GLM best performer) Accuracy: 1.00, precision: 1.00, recall (sensitivity): 1.00, F1-score: 1.00, ROC-AUC: 1.00 for both models (Project 1: malignant vs benign; Project 2: fibrous dysplasia vs ossifying fibroma) AUC: 0.70–0.98 (ML-driven biomarker scoring models), sensitivity: 80–100%, specificity: 80–95% for top models; saliva-based models achieved sensitivity: 95–100% and specificity: 95%; single- microorganism models showed sensitivity: up to 95% but specificity: ≤ 42%, AUC: ≤ 0.56 H2O-AutoML GLM using clinical ± MDCT fascial- space data achieved AUROC ≥ 0.83 for predicting ICU admission in odontogenic infection.

  Clinician-accessible automated machine learning in oral healthcare: A systematic review · 2026 · DOI

• Gomez-Rios et al. [35] Hamdan et al. [25] Kong et al. [26] Kwack et al. [27] F1-score (class of interest – second sedation): 0.83, balanced accuracy: 0.81, precision: 0.92, recall: 0.75; AUC (ROC): 0.92; class “no sedation”—precision: 0.87, recall: 0.96, F1-score: 0.92 Sensitivity (pixel): 86.64%, specificity (pixel): 99.78%, accuracy (pixel): 99.63%, precision (pixel): 82.4%, F1- score (pixel): 0.844; sensitivity (tooth): 98.34%, specificity (tooth): 98.13%, accuracy (tooth): 98.21%, precision (tooth): 98.85%, F1- score (tooth): 0.98; AUC: 0.978 (95% CI 0.959–0.998) Accuracy: 0.981, precision: 0.963, recall: 0.961, specificity: 0.985, F1-score: 0.962; average precision (AUPRC): 0.983; per-class accuracy ranged from 0.965 to 1.000 AUC (training): 0.8283, sensitivity (training): 83.7%, specificity (training): 71.3%; AUC (test): 0.7526, sensitivity (test): 88.6%, specificity (test): 52.8% Cheong et al. [28] Sensitivity: 91.3%, specificity (precision): 92.8%, accuracy: 92%; TP: 63, TN: 64, FP: 5, FN: 6 (Vertex AI image classification model) Gonzalez et al. [29] Sensitivity: 0.888, specificity: 0.988, precision: 0.958, accuracy: 0.964, F1-score: 0.921, ROC-AUC: 0.965 (no- code AI, LandingLens™) variables as the most influential predictors. ORIENTATE AutoML enabled non-technical researchers to build interpretable models; an 8-predictor decision tree achieved F1 = 0.83 and AUC 0.92 for sedation prediction. LandingLens no-code vision model accurately segmented dental restorations (AUC 0.978) with pixel- and tooth- level performance comparable to coded deep-learning systems.

  Clinician-accessible automated machine learning in oral healthcare: A systematic review · 2026 · DOI

• In this context, future research should focus on testing the model with diverse and large datasets and on conducting evaluations without providing anatomical guidance, in order to more clearly elucidate the true visual perception and diagnostic capabilities of GPT-based models. Therefore, additional studies are needed before these results can be generalized to clinical practice.

  Performance of Chat-GPT 5.1 in the Diagnostic Evaluation of Apical Lesions on Panoramic Radiographs · 2026 · DOI

• underscore the need for further research aimed at establishing standardized validation protocols and enhancing algorithmic transparency to facilitate the reliable integration of fuzzy logic AI into orthodontic practice. circumstances.

  Fuzzy Logic as a Bridge to Human-like Artificial Intelligence in Orthodontics: A New Perspective · 2026 · DOI

• Several limitations of this study should be acknowl- edged. Another limitation of this study is the equal weighting of all scoring domains within the composite clinical deci- sion accuracy score, which sums multiple decision-making domains, including systemic risk assessment, treatment planning, medication-related recommendations, and the indication for medical consultation. Finally, the evaluation was limited to a small number of AI models, and rapid model updates or architectural changes may affect performance in future applications.

  Clinical decision accuracy in endodontic treatment of patients with systemic diseases: a comparative analysis using different artificial intelligence models · 2026 · DOI

• Figure 2. Leading Enhancement Assistive Planning research project workflow for deep learning–based orthodontic treatment. STL, standard tessellation language; 3D, three-dimensional; CNN, convolutional neural network; CAD, computeraided design. Figure 3. The Leading Enhancement Assistive Planning system pipeline for orthodontic treatment. 3D, three-dimensional; STL, standard tessellation language; CNN, convolutional neural network; CAD, computer-aided design. https://doi.org/10.4041/kjod25.214 3 Lee et al • LEAP: 3D AI-based tooth classification systemwww.e-kjo.org dental models were collected, representing a balanced distribution of Class I, Class II Division 1, Class II Division 2, Class III, crossbite, deep bite, open bite, and scissor bite malocclusions, as clinically diagnosed by 4 certified orthodontists, each with over 10 years of clinical experience. Figure 4 shows an example of multiview images of an intraoral scan file. The multiview images of the STL files provide a more complete representation, which helps in understanding the 3D structures of the teeth and their supporting structures for machine learning tasks. Using a dataset of 841 STL files, each representing a 3D intraoral model labeled and cross-validated by experts, the system classified cases of deep bite based on learned patterns of malocclusion. Instead of manually defining anatomical landmarks (the lowest point on the mandibular midline), the model used deep learning to automatically extract features and classify malocclusion types. Figure 5 shows a sample visualization of dental arch mesh editing using Autodesk Meshmixer software 3.5.474 (Autodesk, Inc., San Francisco, CA, USA). Autodesk Meshmixer is a free 3D modeling and editing software designed primarily for working with triangular meshes (STL and OBJ files). Table 1 shows the malocclusion types and numbers of samples. The dataset used for model training and evaluation consisted of 841 3D intraoral scan files in STL format. Each scan was labeled by expert orthodontists and cross-validated to ensure annotation reliability. The dataset was balanced across eight clinically relevant malocclusion categories. • Classes I to III: anteroposterior relationships based on Angle’s classification • Malocclusion: deep bite, open bite, crossbite, and Table 1.

  Development of Leading Enhancement Assistive Planning: A three-dimensional tooth classification system for orthodontic treatment with clear aligners · 2026 · DOI

• While the LEAP system has demonstrated accurate classification of malocclusions that can support treat- ment planning, it has a few limitations.23 Current mod- els depend on expert-annotated dataset labels, which, although reviewed and validated by human experts such as dentists and dental technicians, are still likely to contain some subjective bias due to human judgment.24 Additionally, voxel representations effectively preserve spatial information, which is valuable for AI models to understand the relationships between teeth; however, they can be computationally expensive at high resolu- tion. To address this trade-off (high spatial accuracy vs. high computational load), future versions of LEAP may need to explore point clouds (sets of points in space) or mesh-based (surfaces composed of triangles) models to improve performance and efficiency. Furthermore, the system has only been tested on controlled datasets, and its generalizability to different scanners has not yet been validated.

  Development of Leading Enhancement Assistive Planning: A three-dimensional tooth classification system for orthodontic treatment with clear aligners · 2026 · DOI

• LEAP was designed to evolve toward seamless integration into clinical CAD workflows, ultimately serving as a key component of AI-assisted orthodontics. To accommodate diverse clinical environments, we envision multiple deployment strategies, including embedding LEAP as an on-premises module within CAD software, deploying it as a secure cloud-based Application Programming Interface, or connecting through workflow automation tools. These approaches would enable automated classification and treatment planning directly within existing digital systems. However, a critical limitation is that LEAP’s generalizability across different intraoral scanners has not yet been validated, posing a major barrier to clinical applicability. To address this, we plan to conduct a multicenter validation study across diverse clinical settings and scanning devices to enhance system robustness and reliability prior to commercial integration. Integrating LEAP with treatment simulation tools, including predictive models of tooth movements under different aligner protocols, could convert it into a comprehensive treatment planning assistant. Embedding LEAP directly into intraoral scanning devices would enable realtime diagnostic feedback during patient examinations, streamlining workflows and improving chairside efficiency. Ultimately, these advancements position LEAP to improve diagnostic accuracy, reduce clinical workload, and enable more personalized patient care.

  Development of Leading Enhancement Assistive Planning: A three-dimensional tooth classification system for orthodontic treatment with clear aligners · 2026 · DOI

• The substantial heterogeneity in dataset sizes (ranging from approximately 300 to over 10,000 images), variability in ground truth determination methods, and inconsistency in reported performance metrics prevented a reliable quantita- tive comparison across studies. These methodological dis- crepancies precluded the conduct of a formal meta-analysis and limited the ability to derive pooled diagnostic accuracy estimates. Furthermore, our methodological appraisal high- lighted inherent limitations in the risk-of-bias assessment process itself. The traditional QUADAS-2 tool lacks AI- specific criteria and employs a crude categorization system (lacking an intermediate option), which occasionally led to variable inter-rater agreement and challenges in perfectly stratifying the methodological quality of the included AI studies.

  Diagnostic performance of artificial intelligence in periapical radiography: a systematic review · 2026 · DOI

• The strict case-wise stratification ensures no patient overlap between training and testing sets, but inter-observer and intra-observer agreement metrics for the manual CBCT annotations are not reported. No validation is conducted on whether annotators consistently identify edentulous regions, which is critical for assessing ground truth quality in a dataset of 3,755 total annotations.

  Adults’ dental cone beam computed tomography images dataset for detecting and classifying missing teeth · 2026 · DOI

• No analysis is provided on the impact of metal artifacts from dental hardware (restorations, implants, bridges) on detection performance, despite explicitly mentioning metallic artifact challenges in Figure 1. The paper does not evaluate whether metal artifact reduction techniques or artifact-aware training strategies could improve YOLO detector robustness on patients with extensive intra-oral fixed appliances.

  Adults’ dental cone beam computed tomography images dataset for detecting and classifying missing teeth · 2026 · DOI

• The dataset annotations are limited to edentulous regions (missing tooth locations) without classification into specific types of tooth loss (e.g., congenitally missing, extraction, severe periodontal disease, trauma). The paper does not explore whether YOLO-based detectors can simultaneously detect AND differentiate between missing tooth etiologies, which would enhance clinical utility for treatment planning.

  Adults’ dental cone beam computed tomography images dataset for detecting and classifying missing teeth · 2026 · DOI

• The comparative benchmarking excludes recent specialized dental detection models and transformer-based architectures that have shown promise in medical image analysis. The paper does not compare against domain-specific dental AI systems or state-of-the-art foundation models like Segment Anything (referenced in Figure 1) when adapted specifically for missing tooth classification in CBCT data.

  Adults’ dental cone beam computed tomography images dataset for detecting and classifying missing teeth · 2026 · DOI

• The study uses only 2D axial slice-based detection from CBCT scans, despite the inherently 3D nature of volumetric dental imaging. No investigation is conducted on 3D convolutional neural networks or volumetric segmentation approaches for missing tooth detection that could leverage spatial continuity across CBCT slices and potentially improve detection sensitivity.

  Adults’ dental cone beam computed tomography images dataset for detecting and classifying missing teeth · 2026 · DOI

• The dataset contains only 158 patient cases (4,194 2D axial slices) from unspecified CBCT scanners with variable acquisition parameters. The paper does not evaluate how detector performance varies across different CBCT manufacturers, voxel sizes, tube voltages (kV), or exposure settings—factors known to influence image artifacts and detection reliability in dental cone-beam imaging.

  Adults’ dental cone beam computed tomography images dataset for detecting and classifying missing teeth · 2026 · DOI

• The YOLO-based detectors achieved notably low [email protected] scores (14.36-17.20%), indicating poor performance at stricter intersection-over-union thresholds critical for clinical precision in missing teeth detection from CBCT images. The paper does not investigate architectural modifications or training strategies to improve localization accuracy beyond 0.5 IoU threshold, which is essential for reliable dental implant planning applications.

  Adults’ dental cone beam computed tomography images dataset for detecting and classifying missing teeth · 2026 · DOI

• Promising directions include formation of large multicenter databases, validation in prospective clinical studies, and development of ethical-legal mechanisms for safe AI implementation in pediatric dentistry practice.

  Application of artificial intelligence in pediatric dentistry: a systematic review · 2026 · DOI

• There is no clear answer to the question of responsibility for incorrect diagnosis and treatment failure: who is liable - the physician, the AI developer, or both?

  Application of artificial intelligence in pediatric dentistry: a systematic review · 2026 · DOI

Most-cited papers in Dental Radiography and Imaging

• Cone beam computed tomography in <scp>E</scp>ndodontics – a review · International Endodontic Journal · 2014 · 364 citations
• Artificial intelligence in dental research: Checklist for authors, reviewers, readers · Journal of Dentistry · 2021 · 331 citations
• New dimensions in endodontic imaging: part 1. Conventional and alternative radiographic systems · International Endodontic Journal · 2009 · 287 citations
• Deep learning for caries detection: A systematic review · Journal of Dentistry · 2022 · 265 citations
• Diagnosis of Vertical Root Fractures by Cone-beam Computed Tomography in Root-filled Teeth with Confirmation by Direct Visualization: A Systematic Review and Meta-Analysis · Journal of Endodontics · 2021 · 252 citations
• Accuracy of three-dimensional measurements using cone-beam CT · Dentomaxillofacial Radiology · 2006 · 227 citations
• Applications of artificial intelligence in dentistry: A comprehensive review · Journal of Esthetic and Restorative Dentistry · 2021 · 223 citations
• Artificial Intelligence in Endodontics: Current Applications and Future Directions · Journal of Endodontics · 2021 · 200 citations
• A guide for maximizing the accuracy of intraoral digital scans. Part 1: Operator factors · Journal of Esthetic and Restorative Dentistry · 2022 · 196 citations
• Artificial Intelligence for Fast and Accurate 3-Dimensional Tooth Segmentation on Cone-beam Computed Tomography · Journal of Endodontics · 2021 · 178 citations

Most recent work

• Applications for the YOLO deep learning framework in dentistry: A narrative review · Journal of Prosthodontic Research · 2026
• Artificial Intelligence–Driven Dentistry: A Systematic Review of Ethical and Legal Challenges · International Journal of Dentistry · 2026
• Clinician-accessible automated machine learning in oral healthcare: A systematic review · Japanese Dental Science Review · 2026
• Artificial Intelligence in Periodontology: A Systematic Review · Journal of Periodontal Research · 2026
• Acquisition of and quality assessment in a dataset for dental microwear texture analysis (DMTA) of sauropod teeth · Historical Biology · 2026
• The Use of Artificial Intelligence in Planning Dental Implant Procedures: A Systematic Review · Dentistry Journal · 2026
• Knowledge and Practices of Digital Esthetic and Restorative Dentistry among Dental Professionals in Southern Saudi Arabia: A Cross-sectional Survey · World Journal of Dentistry · 2026
• DIAGNOSTIC PERFORMANCE OF MACHINE LEARNING-AIDED PROXIMAL CARIES DETECTION: A SYSTEMATIC REVIEW AND META-ANALYSIS · Journal of Evidence Based Dental Practice · 2026
• Clinical decision accuracy in endodontic treatment of patients with systemic diseases: a comparative analysis using different artificial intelligence models · Odontology · 2026
• Teeth identification and numbering in mixed dentition: evaluating deep learning models for pediatric panoramic radiographs · BMC Oral Health · 2026

Related topics in Dentistry