Espada-Salgado_et

ABSTRACT: This scoping review aimed to map the extent, nature, and characteristics of the available evidence on artificial intelligence (AI) applied to cone-beam computed tomography (CBCT)-based periodontal and peri-implant diagnosis. A scoping review was conducted in accordance with Joanna Briggs Institute methodological guidance and reported following PRISMA-ScR. A systematic search of PubMed/MEDLINE, Scopus, Web of Science, and LILACS/BVS was performed up to April 3, 2026, without language restrictions. Records were exported in RIS format, managed in Zotero for duplicate removal, and screened in Rayyan using predefined eligibility criteria. Original studies applying AI-based methods to CBCT images for periodontal and/or peri-implant diagnostic purposes were included. Data were charted descriptively to summarize study characteristics, diagnostic targets, AI models, performance metrics, and methodological limitations. Of the 358 records identified, 174 were screened after removal of duplicate records and conference papers. Twelve reports underwent full-text assessment, and 7 studies were included in the final qualitative synthesis. Most studies were published between 2024 and 2026 and focused predominantly on periodontal applications, whereas only one addressed peri-implant marginal bone loss. The most consistent findings were observed in narrowly defined tasks, particularly furcation involvement detection, periodontal bone topography segmentation, periodontal defect mapping, and peri-implant marginal bone loss grading. Broader diagnostic platforms showed more variable performance. AI applied to CBCT-based periodontal and peri-implant diagnosis represents an emerging field with still limited methodological development. This review maps the CBCT-specific evidence base and identifies priorities for advancing toward more interpretable, externally validated, and clinically useful AI systems.

KEYWORDS: Artificial intelligence; Cone-beam computed tomography; Periodontitis; Peri-implantitis; Deep learning; Periodontal bone loss.

RESUMEN: Este scoping review tuvo como objetivo mapear la extensión, naturaleza y características de la evidencia disponible sobre la inteligencia artificial (IA) aplicada al diagnóstico periodontal y periimplantario basado en tomografía computarizada de haz cónico (CBCT). Se realizó una scoping review de acuerdo con la guía metodológica del Joanna Briggs Institute y se reportó conforme a PRISMA-ScR. Se efectuó una búsqueda sistemática en PubMed/MEDLINE, Scopus, Web of Science y LILACS/BVS hasta el 3 de abril de 2026, sin restricción de idioma. Los registros se exportaron en formato RIS, se gestionaron en Zotero para la eliminación de duplicados y se cribaron en Rayyan con criterios de elegibilidad predefinidos. Se incluyeron estudios originales que aplicaron métodos de IA a imágenes CBCT con fines diagnósticos periodontales y/o periimplantarios. Los datos se sintetizaron de manera descriptiva. De los 358 registros identificados, 174 fueron cribados tras eliminar duplicados y conference papers. Doce artículos fueron evaluados a texto completo y 7 se incluyeron en la síntesis cualitativa final. La mayoría se publicó entre 2024 y 2026 y se centró predominantemente en aplicaciones periodontales, mientras que solo uno abordó la pérdida ósea marginal periimplantaria. Los hallazgos más consistentes se observaron en tareas diagnósticas específicas, especialmente detección de compromiso de furca, segmentación de la topografía ósea periodontal, mapeo de defectos periodontales y gradación de pérdida ósea marginal periimplantaria. Las plataformas diagnósticas más amplias mostraron un desempeño más variable. La IA aplicada al diagnóstico periodontal y periimplantario basado en CBCT constituye un campo emergente con desarrollo metodológico aún limitado. Esta revisión mapea la evidencia específica en CBCT e identifica prioridades para avanzar hacia sistemas de IA más interpretables, validados externamente y clínicamente útiles.

PALABRAS CLAVE: Inteligencia artificial; Tomografía computarizada de haz cónico; Periodontitis; Periimplantitis; Aprendizaje profundo; Pérdida ósea periodontal.

Introduction

Periodontal and peri-implant diseases remain highly relevant in oral healthcare because their diagnosis depends on the accurate assessment of hard-tissue destruction, defect morphology, and longitudinal bone changes. Although clinical examination is indispensable, conventional diagnostic pathways are still limited by operator variability and by the restricted spatial information provided by two-dimensional imaging. In this context, radiographic interpretation plays a central role in diagnosis, prognosis, and treatment planning, while cone-beam computed tomography (CBCT) has become increasingly relevant in selected periodontal and implant scenarios because it enables three-dimensional visualization of alveolar bone morphology and defect configuration with greater geometric fidelity than conventional radiography (1,2).

At the same time, artificial intelligence (AI) has expanded rapidly across dental imaging and is increasingly being explored in periodontology and implant dentistry. Recent reviews have highlighted its potential for image-based diagnosis, pattern recognition, and decision support, but they have also emphasized substantial heterogeneity in target conditions, imaging sources, analytical approaches, and validation strategies. A major limitation of the current secondary literature is that evidence from periapical, panoramic, photographic, intraoral scanning, and CBCT datasets is often analyzed together, making it difficult to identify the specific contribution of CBCT-based AI to periodontal and peri-implant diagnosis (3-5).

A more focused body of CBCT-based evidence is now emerging. Primary studies have evaluated AI systems against human readers for the detection of periodontal lesions in CBCT datasets and have explored deep learning models for identifying furcation involvement in mandibular molars (6,7). More recent investigations have extended this field to the automated recognition of periodontal bone-defect patterns, segmentation of periodontal bone topography, classification of periodontal disease status, and detection or grading of peri-implant marginal bone loss (8-11).

Despite this progress, the available evidence remains dispersed across different diagnostic targets, datasets, study designs, and model outputs. Therefore, a scoping review is warranted to map the extent, nature, and characteristics of the evidence on artificial intelligence applied to CBCT-based periodontal and peri-implant diagnosis, and to identify current methodological trends, main diagnostic applications, and persisting knowledge gaps relevant to future research and clinical translation.

Materials and Methods

Protocol and registration

This scoping review was conducted in accordance with the updated methodological guidance of the Joanna Briggs Institute for scoping reviews (12) and was reported following the PRISMA Extension for Scoping Reviews (PRISMA-ScR) (13). To enhance methodological transparency, the review protocol and supporting materials were registered in the Open Science Framework (OSF) Registries. In addition, the associated OSF project was used as a repository for the protocol, search strategies, screening records, and supplementary review materials. The associated project is available at: https://osf.io/dzvkh

Eligibility criteria

The review question was structured using the Population–Concept–Context (PCC) framework. The population comprised studies involving human participants, human CBCT datasets, or CBCT image datasets representing periodontal tissues, alveolar bone support, teeth with periodontal involvement, dental implants, or peri-implant tissues. The concept of interest was the application of artificial intelligence methods, including machine learning, deep learning, neural networks, convolutional neural networks, computer vision, radiomics, and related automated analytical approaches, to CBCT-based periodontal and peri-implant diagnosis. The context included clinical, academic, laboratory, or research settings in which CBCT was used to assess periodontal and/or peri-implant conditions.

Studies were considered eligible when they were original investigations applying AI-based methods to CBCT images for the evaluation of periodontal and/or peri-implant conditions. Eligible applications included detection, diagnosis, classification, segmentation, quantification, measurement, severity grading, or prognostic assessment directly related to periodontal or peri-implant tissues. Eligible diagnostic targets included periodontal bone loss, alveolar bone loss, intra bone defects, furcation involvement, periodontal defect morphology, peri-implant bone loss, peri-implantitis, peri-implant defect morphology, or related CBCT-based diagnostic parameters. No language restrictions were applied. The publication period extended from January 1, 2015 to April 3, 2026, corresponding to the date of the final search update.

Reviews, systematic reviews, scoping reviews, bibliometric studies, editorials, commentaries, letters, notes, protocols, book chapters, dissertations, theses, animal studies, studies not focused on CBCT, studies unrelated to periodontal or peri-implant diagnosis, and studies without an AI-based analytical component were excluded. Conference papers were also excluded before formal screening to maintain consistency with the final eligibility framework.

Information sources and search strategy

A systematic literature search was conducted in four electronic databases: PubMed/MEDLINE, Scopus, Web of Science, and the Virtual Health Library (LILACS/BVS). The final search update was performed on April 3, 2026. Search strategies combined controlled vocabulary, when applicable, and free-text terms related to three core concepts: artificial intelligence, cone-beam computed tomography, and periodontal/peri-implant conditions. MeSH terms were used in PubMed, and DeCS terms were incorporated into the LILACS/BVS strategy. The syntax for Scopus and Web of Science was adapted to their specific search fields and indexing systems.

The search terms covered AI-related concepts such as artificial intelligence, machine learning, deep learning, neural networks, convolutional neural networks, computer vision, and radiomics; imaging-related terms such as cone-beam computed tomography and CBCT; and target-condition terms such as periodontitis, peri-implantitis, periodontal bone loss, alveolar bone loss, intra bone defects, furcation involvement, and peri-implant bone loss. Full search strings for all databases are available in the supplementary material and in the associated OSF project.

Selection of sources of evidence

All records retrieved from the databases were exported in RIS format and imported into Zotero for reference management and duplicate removal. After deduplication, conference papers were removed prior to formal screening. The remaining records were then uploaded to Rayyan, which was used to support title and abstract screening and to facilitate the preselection of potentially relevant studies (14).

Title and abstract screening was performed independently by two reviewers. Each reviewer classified records as eligible for full-text assessment or excluded according to the predefined PCC framework and eligibility criteria. Disagreements between the two reviewers were resolved through discussion; when consensus could not be reached, a third reviewer acted as adjudicator and made the final decision. Inter-rater agreement before consensus was assessed using Cohen’s kappa.

Potentially eligible reports underwent full-text assessment independently by two reviewers against the predefined PCC framework and eligibility criteria. Disagreements at this stage were resolved through consensus discussion, with a third reviewer acting as adjudicator when required. Reasons for exclusion at the full-text stage were documented and incorporated into the PRISMA flow diagram. The overall study selection process was summarized according to PRISMA-ScR recommendations.

Data charting process

Data from the included studies were charted using a predefined extraction framework developed specifically for this review. After full-text assessment, the selected articles were examined in detail and their relevant characteristics were recorded in a structured spreadsheet designed to support descriptive evidence mapping. Because the aim of a scoping review is to characterize the available literature rather than to generate pooled effect estimates, the charting process was oriented toward identifying and organizing the main methodological and diagnostic features of the included studies.

Data items

For each included study, the following variables were considered during data charting: first author, year of publication, country or setting, study design, sample or dataset characteristics, periodontal or peri-implant diagnostic target, type of CBCT input, AI approach or model, main analytical task, reference standard when reported, performance metrics, and principal findings. Additional study-specific details relevant to interpretation, such as segmentation targets, grading categories, comparison with human readers, or disease classification outputs, were also recorded when available.

Synthesis of results

The extracted evidence was synthesized descriptively. The results were organized according to the main diagnostic target and the type of AI application, with emphasis on periodontal versus peri-implant use, lesion detection, defect classification, segmentation-based assessment, and automated grading approaches. Given the methodological heterogeneity across the included studies in terms of design, CBCT datasets, analytical strategies, and reported outcomes, no meta-analysis was planned or conducted. Instead, the review aimed to map the extent, nature, and characteristics of the evidence and to identify current methodological trends and knowledge gaps relevant to future research and clinical translation.

RESULTS

Selection and general characteristics of the included studies

After removal of duplicate records and conference papers, 174 records were screened by title and abstract. Title and abstract screening were performed independently by two reviewers using Rayyan. The reviewers agreed on 170 records and disagreed on 4 records, resulting in an observed agreement of 97.7%. Cohen’s kappa was 0.81, indicating substantial-to-almost-perfect agreement. Disagreements were resolved through discussion, with a third reviewer acting as adjudicator when required. The agreement matrix is presented in Supplementary Table S1.

Following title and abstract screening, 12 reports were sought for retrieval and assessed for eligibility at full-text level. Of these, 7 studies were included in the final qualitative synthesis and 5 reports were excluded with reasons. The study selection process is presented in Figure 1. The main characteristics of the included studies are summarized in Table 1.

The included studies were published between 2021 and 2026, with a clear increase in publications from 2024 onward. Of the 7 studies, 6 focused primarily on periodontal applications and 1 addressed peri-implant marginal bone loss. The evidence was geographically heterogeneous, including studies from Europe, the Middle East, and multicenter or international collaborations, but no included study originated from Latin America. Most studies were retrospective diagnostic or technical validation investigations, although their designs varied substantially, ranging from proof-of-concept classification studies and segmentation pipelines to observer-versus-AI comparisons and clinician-aided diagnostic evaluations.

Performance of the included artificial intelligence applications

The main performance metrics and key findings of the included studies are presented in Table 2. Three studies reported high or strong performance in narrowly defined diagnostic tasks. Palkovics et al. evaluated a multi-stage SegResNet-based model for automatic segmentation of mandibular periodontal bone topography in advanced periodontitis and reported a mean Dice similarity coefficient of 0.9650, an IoU of 0.9340, and an HD95 of 0.482 mm. The model also showed high agreement with semi-automatic measurements on mesial, distal, and lingual surfaces, while reducing segmentation time by 47-fold. Shetty et al. reported strong performance for binary detection of furcation involvement in mandibular first molars using ResNet101V2 with transfer learning, with a test accuracy of 91%, precision of 0.98, F1-score of 0.98, and AUC of 0.98. Similarly, Madani and Bashizadeh Fakhar found that a YOLOv8-based model achieved an overall accuracy, precision, recall, and F1-score of 0.90 for peri-implant marginal bone loss detection and grading, with mAP@0.5 of 0.889 and excellent performance for healthy and mild lesions, although performance declined in moderate and severe categories.

Other studies showed more moderate or exploratory results. Burlea et al. developed a 3D ResNet-18 classifier for periodontitis versus healthy status using 403 CBCT volumes and reported an AUROC of 0.729, AUPRC of 0.551, and overall accuracy of 0.62 at a high-sensitivity operating point. Despite the modest discriminative performance, Grad-CAM maps consistently highlighted clinically relevant anatomical regions, particularly the alveolar crest and furcation areas. Kurt-Bayrakdar et al. proposed a multi-stage framework that combined tooth segmentation, periodontal defect segmentation, and healthy/unhealthy periodontal classification. Their results showed strong performance for total alveolar bone loss and perio-endo lesions, but substantially lower AUC values for supra-bony, infra-bony, buccal, and furcation defects before image cropping. Cropping improved several of these values, but the variability across targets remained considerable. The final healthy/unhealthy classifier achieved an accuracy of 80% in healthy individuals and 76% in unhealthy individuals.

In contrast, Schulze et al. reported that the evaluated commercial CNN underperformed relative to human observers in the detection of periodontal lesions in CBCT datasets. Human readers achieved good to very good agreement for most lesion categories, whereas the AI system showed significantly lower performance and failed to detect any periradicular osteolysis. This study is important because it highlights that not all commercially available CBCT-AI tools are necessarily reliable for periodontal lesion detection, even when performance appears promising in other dental applications.

The study by Ezhov et al. occupied an intermediate position within the evidence base. Rather than focusing exclusively on periodontology, it evaluated a broad CBCT-based diagnostic platform that included a periodontitis module among several other tasks. The overall AI system showed high sensitivity and specificity, and the AI-aided group of clinicians demonstrated higher overall sensitivity and shorter interpretation times than the unaided group. However, periodontal findings were embedded within a broader diagnostic framework, and the study did not isolate the periodontal component with the same granularity as the more targeted periodontal or peri-implant investigations.

Coverage by diagnostic focus and artificial intelligence task

The distribution of included studies according to diagnostic target and application area is shown in Figure 2. The evidence was concentrated mainly on periodontal targets, especially periodontal bone loss, furcation involvement, periodontal bone-defect patterns, and periodontal disease classification. Only one included study specifically addressed peri-implant bone loss. In methodological terms, the reviewed studies were distributed across three main AI application types: segmentation-based approaches, detection/classification models, and broader decision-support systems with periodontal modules.

Segmentation-based applications were represented most clearly by Palkovics et al. and Kurt-Bayrakdar et al., both of which used CBCT to segment periodontal structures or periodontal defect patterns. Detection or classification tasks were addressed by Shetty et al., Burlea et al., Madani and Bashizadeh Fakhar, and Schulze et al., whereas Ezhov et al. evaluated a broader CBCT-based diagnostic platform that included periodontitis among several diagnostic modules. Overall, the evidence suggests that CBCT-based AI in this field has developed primarily around structural interpretation of alveolar bone loss and defect morphology, rather than around multimodal or prognostic modeling.

Cross-study methodological patterns

and limitations

The main limitations reported by the included studies and their implications for evidence interpretation are summarized in Table 3. Across the included studies, several recurrent methodological patterns were identified. First, most investigations were based on retrospective or archive-derived datasets, often from a single source or limited number of centers. Second, the majority of applications focused on imaging-only tasks, with little integration of clinical or longitudinal variables. Third, performance appeared to depend strongly on the complexity of the target condition. Broader or more visually evident targets such as total alveolar bone loss, overall periodontal status, or healthy versus mild peri-implant bone loss tended to show better performance than more localized or morphologically subtle defects.

The limitations summarized in Table 3 indicate consistent barriers to clinical translation. Small test sets, single-source data, lack of external validation, class imbalance, exclusion of high-artifact scans, and variable robustness across anatomical regions were common. In segmentation studies, performance dropped in more difficult surfaces or smaller localized defects. In classification studies, moderate performance metrics suggested proof-of-concept feasibility rather than immediate clinical readiness. In comparative studies, commercial AI tools did not always meet the level of expert human readers for periodontal lesion detection. Taken together, these findings indicate that the evidence base is promising but still methodologically limited, with stronger support for targeted proof-of-concept applications than for generalized clinical deployment.

Synthesis of the current evidence

The chronological distribution of the included studies is presented in Figure 3. Most of the evidence was published between 2024 and 2026, suggesting a recent acceleration of interest in CBCT-based AI applications for periodontal and peri-implant diagnosis.

Overall, the included literature suggests that AI applications for CBCT-based periodontal and peri-implant diagnosis are developing rapidly but remain concentrated in a limited number of diagnostic targets. The strongest evidence currently supports segmentation of periodontal bone topography, detection of furcation involvement, and automated grading of peri-implant marginal bone loss. Broader classification frameworks and commercial diagnostic support systems have shown potential, but their performance remains variable and, in some cases, clearly inferior to human assessment for specific periodontal lesions. Taken together, the evidence map and chronological distribution indicate a field in active expansion, with increasing methodological diversity but still limited standardization, restricted external validation, and only sparse peri-implant representation.

Figure 1. PRISMA 2020 flow diagram of study selection.

Table 1. Main characteristics of the included studies.

Study	Year	Diagnostic focus	CBCT dataset	AI model	Main AI task
Palkovics et al.	2025	Advanced periodontitis; periodontal bone topography and defects	70 scans for model development (57 train, 13 validation) and 10 independent test scans	Multi-stage SegResNet-based deep learning model	Automatic segmentation; periodontal-region measurement support
Shetty et al.	2024	Periodontal furcation involvement	285 eligible axial CBCT images; data augmentation to 600 images; 85 images for testing	ResNet101V2 with transfer learning	Binary detection/classification
Madani and Bashizadeh Fakhar	2026	Peri-implant marginal bone loss	699 2D CBCT sections	YOLOv8-based detector	Detection and grading
Burlea et al.	2025	Periodontal disease classification	403 CBCT volumes from 403 patients (282 train, 60 validation, 61 test)	3D ResNet-18 classifier with Grad-CAM	Binary classification with biological concordance analysis
Schulze et al.	2024	Periodontal lesions in CBCT	38 CBCT datasets (up to 1,216 teeth), evaluated by three human readers and a CNN	Diagnocat deep CNN	Automated lesion detection compared with human observers
Ezhov et al.	2021	General dental CBCT diagnosis including a periodontitis module	1346 CBCT scans for training; 300 CBCT scans for performance evaluation; 30 scans evaluated by 24 dentists in aided vs unaided comparison	Multi-module CNN-based Diagnocat system	Automated localization/classification and clinician-aided diagnosis
Kurt-Bayrakdar et al.	2025	Patterns of periodontal bone loss and periodontal status	502 subjects; 250 CBCT volumes for tooth segmentation/numbering; 251 diseased volumes for periodontal defect segmentation; healthy and diseased scans used for classification	nnU-Net v2 models plus TensorFlow/Keras CNN classifier	Segmentation, tooth numbering, and healthy/unhealthy classification

Abbreviations: AI, artificial intelligence; CBCT, cone-beam computed tomography; CNN, convolutional neural network; Grad-CAM, gradient-weighted class activation mapping; nnU-Net, no-new-U-Net; YOLO, You Only Look Once.

Table 2. Main performance metrics and key findings of the included studies.

Study	Main target(s)	Key performance metrics	Interpretation of findings
Palkovics et al.	Mandibular periodontal bone topography	Mean DSC 0.9650, IoU 0.9340, HD95 0.482 mm; ICCs mesial 0.9442, distal 0.9232, lingual 0.9598, buccal 0.7481; 47-fold faster than semi-automatic segmentation	High/strong performance
Shetty et al.	Furcation involvement in mandibular first molars	Training accuracy 98%, validation accuracy 97%, test accuracy 91%, precision 0.98, F1-score 0.98, AUC 0.98, test loss 0.2170	High/strong performance
Madani and Bashizadeh Fakhar	Peri-implant marginal bone loss (healthy, mild, moderate, severe)	Overall accuracy/precision/recall/F1-score 0.90; healthy F1 0.95; mild F1 0.90; moderate F1 0.62; severe F1 0.70; three-class accuracy 0.88; Kappa 0.954; mAP@0.5 0.889; recall 0.98; implant-length detection accuracy 1.00	High/strong performance
Burlea et al.	Periodontitis vs healthy status	AUROC 0.729 (95% CI 0.599–0.850), AUPRC 0.551 (95% CI 0.404–0.727), high-sensitivity operating point sensitivity 0.95 with specificity 0.48, overall accuracy 0.62; Grad-CAM highlighted alveolar crest and furcation regions	Moderate/exploratory performance
Schulze et al.	Vertical bone loss, furcation involvement, periradicular osteolysis	CNN performance was significantly lower than human readers; Fleiss kappa among observers was 0.43 for vertical lesions, 0.80 for initial furcation, 0.88 for severe furcation, and 0.94 for periradicular osteolysis; the CNN failed to detect any periradicular osteolysis	Inferior to human readers
Ezhov et al.	Alveolar bone loss/periodontitis as one of multiple diagnostic modules	AI overall sensitivity 0.9239 and specificity 0.9899; aided vs unaided overall sensitivity 0.8537 vs 0.7672 and specificity 0.9672 vs 0.9616; aided group had shorter evaluation time (17.55 vs 18.74 min, p=0.032)	Moderate/exploratory performance
Kurt-Bayrakdar et al.	Total alveolar bone loss, supra-bony defects, infra-bony defects, perio-endo lesions, buccal defects, furcation defects	Tooth segmentation AUC 0.9594 (accuracy 0.99); periodontal defect AUCs: total alveolar bone loss 0.8499, supra-bony 0.5052, infra-bony 0.5613 (cropped 0.7488), perio-endo 0.8893, buccal 0.6780 (cropped 0.7592), furcation 0.6332 (cropped 0.8087); classifier accuracy 80% in healthy and 76% in unhealthy individuals	Moderate/exploratory performance

Abbreviations: AUC, area under the curve; AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; CNN, convolutional neural network; DSC, Dice similarity coefficient; F1-score, harmonic mean of precision and recall; Grad-CAM, gradient-weighted class activation mapping; HD95, 95th percentile Hausdorff distance; ICC, intraclass correlation coefficient; IoU, intersection over union; mAP, mean average precision.

Note: Performance interpretation was synthesized descriptively according to the metrics and conclusions reported in each included study.

Figure 2. Evidence map of included studies according to diagnostic target and application area.

Table 3. Main limitations reported by the included studies and their implications for evidence interpretation.

Study	Main limitations reported	Implication for evidence synthesis
Palkovics et al.	Small test set; lower consistency at buccal surface; robustness requires improvement	Suggests that segmentation of advanced periodontal topography is feasible, but stability across surfaces and broader generalizability remain uncertain.
Shetty et al.	Preliminary study; relatively small dataset; single target condition	Supports furcation detection as a promising target, but the evidence remains preliminary and requires external validation on larger datasets.
Madani and Bashizadeh Fakhar	Reduced performance in moderate/severe cases; class imbalance; excluded high-artifact scans; external validation required	Shows promise for peri-implant grading, but class imbalance and artifact-sensitive exclusions may limit real-world applicability.
Burlea et al.	Moderate discriminative performance; single-source imaging; exploratory omics triangulation rather than subject-level validation	Indicates proof-of-concept feasibility for CBCT-based periodontitis classification, but present evidence is exploratory rather than clinically definitive.
Schulze et al.	Small archive-based sample; AI underperformed for periodontal lesions; limited clinical utility in current form	Highlights that current commercial AI solutions may underperform in periodontal lesion detection despite acceptable human agreement.
Ezhov et al.	Periodontal findings embedded within a broader dental system; periodontal bone loss sensitivity in aided reads remained low; per-condition detail limited	Suggests broad diagnostic benefit from AI support, but periodontal interpretation cannot be isolated cleanly from other dental modules.
Kurt-Bayrakdar et al.	Performance varied markedly by defect type; localized millimeter-scale defects required cropping; external validation not reported	Demonstrates broad diagnostic coverage, although performance heterogeneity across defect types limits uniform clinical applicability.

Figure 3. Chronological distribution of included studies by primary application focus.

DISCUSSION

This scoping review shows that artificial intelligence applied specifically to CBCT-based periodontal and peri-implant diagnosis is no longer a conceptual possibility, but it is not yet a mature clinical field. Its main contribution lies in isolating a CBCT-specific body of evidence from a much broader AI literature that usually merges panoramic, periapical, photographic, intraoral scanning, and tomographic data within the same discussion. By restricting the synthesis to CBCT, the present review demonstrates that the available evidence is smaller, more technically focused, and methodologically less mature than broader reviews in periodontology and implant dentistry may suggest. In practical terms, the field is advancing, but it is still dominated by proof-of-concept studies rather than by externally validated clinical tools ready for routine deployment (3,5).

A second key finding is that the strongest results were reported in narrowly defined diagnostic tasks, not in broad or generalized diagnostic platforms. High performance was observed in studies targeting furcation involvement, segmentation of periodontal bone topography, and grading of peri-implant marginal bone loss, whereas broader disease-level classification and commercial decision-support systems showed more variable results. This distinction is highly relevant for clinicians because current CBCT-based AI seems to perform best when asked to solve a specific anatomical or morphometric problem, but becomes less consistent when expected to provide generalized periodontal diagnostic support across heterogeneous lesions and clinical scenarios. In that sense, the current evidence does not yet support an uncritical transition from AI-assisted image analysis to fully reliable automated diagnosis (6,7). More recent task-oriented studies reinforce this interpretation, as models designed for periodontal defect mapping and bone-topography segmentation achieved their best performance in structured and well-delimited outputs rather than in comprehensive disease interpretation (8,9). Likewise, exploratory CBCT-based classifiers and broader diagnostic support systems have shown potential, but with performance levels and consistency that still fall short of what would be required for generalized clinical autonomy (10,11).

An important implication of this review is that the field is beginning to move beyond simple lesion detection toward digital phenotyping and structured volumetric interpretation. Recent adjacent studies not included in the final synthesis because of scope decisions point in the same direction. AI-driven gingival segmentation on CBCT has shown that soft-tissue contour extraction may become sufficiently accurate to complement hard-tissue interpretation in digital workflows, while technical evaluations of automated radiological reporting systems indicate that CBCT-AI can already perform adequately in selected descriptive tasks but remains inconsistent for periodontal-support assessment (15,16). In parallel, automatic periodontal ligament segmentation and multimodal systems such as PerioAI suggest that the next phase of development may involve integration of CBCT with other digital inputs to derive clinically interpretable periodontal markers, rather than relying only on isolated binary predictions (17,18). This is especially relevant because the real clinical value of AI in periodontology will probably depend less on producing another high-accuracy number and more on generating outputs that clinicians can interpret, verify, and incorporate into treatment planning with confidence.

The review also highlights a persistent methodological problem: technical feasibility is progressing faster than translational validation. Most included studies relied on retrospective, single-source, or highly curated datasets, often with limited external testing and substantial preprocessing. Some models operated on cropped CBCT slices or selected regions rather than on the full complexity of real-world scans, which likely inflates performance relative to routine practice. This concern is consistent with more recent secondary evidence. A systematic review using APPRAISE-AI found that studies on AI for periodontal bone loss detection still suffer from important reporting and validation limitations, while a recent scoping review in periodontology similarly concluded that the field is expanding quickly but remains uneven in methodological rigor and clinical readiness (19,20). For the current literature, the challenge is therefore not merely to demonstrate that AI can work on CBCT images, but to demonstrate that it remains robust across devices, centers, artifact levels, anatomical complexity, and clinically relevant disease definitions.

The peri-implant domain deserves special emphasis because it was clearly underrepresented in the present synthesis. Only one included study directly addressed peri-implant marginal bone loss on CBCT, despite the obvious clinical relevance of peri-implant disease and the known value of three-dimensional imaging for complex peri-implant bone assessment. This imbalance suggests that the periodontal field is currently driving most CBCT-AI innovation, whereas peri-implant applications are still in an earlier phase of development. Broader recent reviews in implant dentistry indicate that AI use is growing rapidly in implant planning, identification, prognosis, and radiographic analysis, and that AI-based radiographic assessment of marginal bone loss may already achieve strong overall diagnostic performance (21,22). However, the available peri-implant evidence remains dominated by non-CBCT or mixed-modality studies. Consequently, one of the clearest outputs of this scoping review is the identification of peri-implant CBCT-AI as a strategically important but still insufficiently developed research area.

This scoping review should also be interpreted in light of some limitations. The number of included studies was limited, and the evidence base was characterized by substantial heterogeneity in diagnostic targets, datasets, AI architectures, analytical tasks, and outcome reporting. Most included investigations were retrospective and proof-of-concept studies, frequently based on single-source data and with limited external validation, which restricts the strength of clinical inferences. In addition, peri-implant applications were markedly underrepresented relative to periodontal applications. Nevertheless, these limitations also reflect the current developmental stage of the field and reinforce the relevance of this review as a mapping exercise of an emerging and still fragmented evidence base.

From a clinical and research perspective, the main message of this review is that the field should now move from accuracy-centered experimentation toward decision-ready systems. Future studies should prioritize multicenter external validation, harmonized periodontal and peri-implant case definitions, full volumetric testing under realistic artifact conditions, calibration analysis, transparent reporting of failure modes, and direct comparison with expert readers in clinically meaningful settings. The development of more advanced AI pipelines for periodontal bone loss diagnosis and the broader implant-dentistry evidence base already point to this transition, but CBCT-specific research still requires stronger standardization and translational discipline (23,24). The real contribution of this scoping review, therefore, is to clarify that the next frontier is not AI in isolation, but clinically interpretable, externally validated, CBCT-based AI capable of supporting periodontal and peri-implant decision-making in real practice.

CONCLUSIONS

A clear message emerges from the current evidence: artificial intelligence applied to CBCT-based periodontal and peri-implant diagnosis has progressed from conceptual promise to targeted proof-of-concept utility. The most consistent advances have been observed in well-defined tasks such as periodontal bone-defect assessment, furcation involvement detection, periodontal bone topography segmentation, and peri-implant marginal bone loss grading, indicating that CBCT-based AI may become a valuable adjunct in complex three-dimensional diagnostic scenarios.

Important gaps remain. The evidence base is still limited, heterogeneous, and dominated by retrospective studies with restricted external validation, while peri-implant applications remain notably underrepresented. The main contribution of this scoping review is to provide a focused map of the CBCT-specific evidence base, clarify where meaningful progress has already been achieved, and identify the methodological priorities needed to move the field toward clinically interpretable, externally validated, and decision-oriented systems for real-world periodontal and peri-implant care.

Funding: None.

Conflicts of Interest: The authors declare no potential conflicts of interest regarding this article's authorship and/or publication.

Author contribution statement: Conceptualization and Study Design, F.M.E.S. and M.M.I.; Literature Search and Screening, F.M.E.S. and M.M.I.; Eligibility Assessment and Study Selection, F.M.E.S. and M.M.I., with S.V.S. acting as the third reviewer for disagreement resolution; Methodology and Validation, F.M.E.S. and M.M.I.; Data Charting, Analysis and Interpretation, F.M.E.S. and M.M.I.; Writing-Original Draft Preparation, F.M.E.S.; Writing-Review and Editing, F.M.E.S., M.M.I., S.V.S., B.O., and R.A.Y.; Additional Literature Search and Writing Support, B.O. and R.A.Y.; Supervision, M.M.I.; Project Administration, F.M.E.S. and M.M.I. All authors read and approved the final manuscript.

REFERENCES

1. Jacobs R., Fontenele R.C., Lahoud P., Shujaat S., Bornstein M.M. Radiographic diagnosis of periodontal diseases: current evidence versus innovations. Periodontol 2000. 2024; 95 (1): 51-69. doi:10.1111/prd.12580.

2. Galarraga-Vinueza M.E., Pagni S., Finkelman M., Schoenbaum T., Chambrone L. Prevalence, incidence, systemic, behavioral, and patient-related risk factors and indicators for peri-implant diseases: an AO/AAP systematic review and meta-analysis. J Periodontol. 2025; 96 (6): 587-633. doi:10.1002/JPER.24-0154

3. Pitchika V., Büttner M., Schwendicke F. Artificial intelligence and personalized diagnostics in periodontology: a narrative review. Periodontol 2000. 2024; 95 (1): 220-231. doi:10.1111/prd.12586

4. Șalgău C.A., Morar A., Zgarta A.D., Ancuța D.L., Rădulescu A., Mitrea I.L., et al. Applications of machine learning in periodontology and implantology: a comprehensive review. Ann Biomed Eng. 2024; 52 (9): 2348-2371. doi:10.1007/s10439-024-03559-0

5. Roy R., Chopra A., Karmakar S., Bhat S.G. Applications of artificial intelligence for diagnosis of periodontal/peri-implant diseases: a narrative review. J Oral Rehabil. 2025; 52 (8): 1193-1219. doi:10.1111/joor.14045

6. Schulze D., Häußermann L., Ripper J., Sottong T. Comparison between observer-based and AI-based reading of CBCT datasets: an interrater-reliability study. Saudi Dent J. 2024; 36 (2): 291-295. doi:10.1016/j.sdentj.2023.11.001

7. Shetty S., Talaat W., AlKawas S., Al-Rawi N., Reddy S., Hamdoon Z., Kheder W., Acharya A., Ozsahin D.U., David L.R. Application of artificial intelligence-based detection of furcation involvement in mandibular first molar using cone beam tomography images: a preliminary study. BMC Oral Health. 2024; 24:1476. doi:10.1186/s12903-024-05268-5

8. Kurt-Bayrakdar S., Bayrakdar I.S., Kuran A., Çelik Ö., Orhan K., Jagtap R. Advancing periodontal diagnosis: harnessing advanced artificial intelligence for patterns of periodontal bone loss in cone-beam computed tomography. Dentomaxillofac Radiol. 2025; 54 (4): 268-278. doi:10.1093/dmfr/twaf011

9. Palkovics D., Molnar B., Pinter C., García-Mato D., Diaz-Pinto A., Windisch P., Ramseier C.A. Automatic deep learning segmentation of mandibular periodontal bone topography on cone-beam computed tomography images. J Dent. 2025; 159: 105813. doi:10.1016/j.jdent.2025.105813

10. Burlea Ș.L., Buzea C.G., Nedeff F., Mirilă D., Nedeff V., Agop M., et al. Deep learning analysis of CBCT images for periodontal disease: phenotype-level concordance with independent transcriptomic and microbiome datasets. Dent J (Basel). 2025; 13 (12): 578. doi:10.3390/dj13120578

11. Madani Z., Bashizadeh Fakhar H. Detection and classification of peri-implant marginal bone loss in cone-beam computed tomography using a deep learning approach. Clin Exp Dent Res. 2026;12: e70308. doi:10.1002/cre2.70308

12. Peters M.D.J., Godfrey C., McInerney P., Munn Z., Tricco A.C., Khalil H. Updated methodological guidance for the conduct of scoping reviews. JBI Evid Synth. 2020; 18 (10): 2119-26.

13. Tricco A.C., Lillie E., Zarin W., O’Brien K.K., Colquhoun H., Levac D., et al. PRISMA Extension for Scoping Reviews (PRISMA-ScR): checklist and explanation. Ann Intern Med. 2018; 169 (7): 467-73.

14. Ouzzani M., Hammady H., Fedorowicz Z., Elmagarmid A. Rayyan—a web and mobile app for systematic reviews. Syst Rev. 2016; 5: 210.

15. Fagundes F.B., Papasratorn D., Elgarba B.M., Fontenele R.C., Neves F.S., Jacobs R. AI-driven gingival segmentation on CBCT: validation using delineation by intraoral scanning and CBCT-based cotton roll separation. J Dent. 2026;166:106331. doi:10.1016/j.jdent.2026.106331

16. Feltraco L.T., Rossetto C., Yeung A.W.K., Soares M.Q.S., Oenning A.C. Utility of the radiological report function of an artificial intelligence system in interpreting CBCT images: a technical report. Dentomaxillofac Radiol. 2025; 54 (3): 239-244. doi:10.1093/dmfr/twaf004

17. Su S., Jia X., Zhan L., Gao S., Zhang Q., Huang X. Automatic tooth periodontal ligament segmentation of cone beam computed tomography based on instance segmentation network. Heliyon. 2024; 10: e24097. doi:10.1016/j.heliyon.2024.e24097

18. Tan M., Cui Z., Li Y., Fang Y., Mei L., Zhao Y., et al. PerioAI: A digital system for periodontal disease diagnosis from an intra-oral scan and cone-beam CT image. Cell Rep Med. 2025; 6 (6): 102186. doi:10.1016/j.xcrm.2025.102186

19. Khubrani Y.H., Thomas D., Slator P.J., White R.D., Farnell D.J.J. Detection of periodontal bone loss and periodontitis from 2D dental radiographs via machine learning and deep learning: systematic review employing APPRAISE-AI and meta-analysis. Dentomaxillofac Radiol. 2025; 54 (2): 89-108. doi:10.1093/dmfr/twae070

20. Chatzopoulos G.S., Koidou V.P., Tsalikis L., Kaklamanos E.G. Clinical applications of artificial intelligence in periodontology: a scoping review. Medicina (Kaunas). 2025; 61 (6): 1066. doi:10.3390/medicina61061066

21. Vázquez-Sebrango G., Anitua E., Macía I., Arganda-Carreras I. The role of artificial intelligence in implant dentistry: a systematic review. Int J Oral Maxillofac Surg. 2025. doi:10.1016/j.ijom.2025.04.005

22. Li Y., Wang X., Zhu H., Ye W. The diagnostic performance of AI based on dental radiographs in predicting marginal bone loss around dental implants: a systematic review and meta-analysis. J Prosthet Dent. 2025; 134 (6): 2190.e1-2190.e11. doi:10.1016/j.prosdent.2025.08.021

23. Rezallah N.N.F., Sherif G., Abdelkarim A.Z., Afifi S. Enhancing periodontal bone loss diagnosis through advanced AI techniques. Appl Sci. 2025; 15 (12): 6832. doi:10.3390/app15126832.

24. Emami M., Shirani M. Application and performance of artificial intelligence in implant dentistry: an umbrella review. Dent Rev. 2025; 5 (3): 100159. doi:10.1016/j.dentre.2025.100159

Supplementary

Table S1. Inter-rater agreement matrix for title and abstract screenin.

	Reviewer 2: Include for full-text assessment	Reviewer 2: Exclude	Total
Reviewer 1: Include for full-text assessment	9	2	11
Reviewer 1: Exclude	2	161	163
Total	11	163	174