Cognitive Computing Approaches for Assessing Mandibular Third Molar Impaction: A Systematic Review

Cognitive Computing Approaches for Assessing Mandibular Third Molar Impaction: A Systematic Review | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Systematic Review Cognitive Computing Approaches for Assessing Mandibular Third Molar Impaction: A Systematic Review Muhammadh Munthasir, Prathibha G, Parimala Sagar, Kavitha Prasad This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8705343/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Context: Mandibular third molar impaction presents significant diagnostic and surgical challenges. Artificial intelligence, particularly deep learning approaches, has emerged as a promising tool for classification, risk assessment, and surgical planning in oral and maxillofacial surgery. Objective: To systematically evaluate the diagnostic performance and clinical applicability of cognitive computing techniques for assessing mandibular third molar impaction and associated surgical complexities. Evidence Acquisition: A systematic literature search was conducted in accordance with PRISMA 2020 guidelines using PubMed, Scopus, Google Scholar Web of Science, and IEEE Xplore databases for studies published up to January 2025. Search terms included combinations of "third molar," "wisdom tooth," "impaction," "artificial intelligence," "deep learning," "radiograph," and "convolutional neural networks." Two reviewers independently screened and extracted data. Risk of bias was assessed using adapted Joanna Briggs Institute (JBI) quality assessment tools. Results: Fifteen studies met inclusion criteria, involving primarily convolutional neural networks (CNNs), support vector machines (SVMs), and ensemble classifiers. Most studies utilised panoramic radiographs, whilst some incorporated cone-beam computed tomography (CBCT) or clinical metadata. AI systems demonstrated diagnostic accuracies ranging from 78.91% to 99.0% for impaction classification, and 72.32% to 99.0% for mandibular canal proximity assessment. Multiple studies showed that AI models outperformed dental students and general practitioners, though performance varied across different architectural approaches and training datasets. Several models demonstrated potential for predicting extraction time (mean absolute error < 3 minutes) and postoperative complications (accuracy up to 98%). Limitations: Limited generalisability due to single-institution datasets, variable dataset sizes, and heterogeneous methodological approaches. Publication bias and selective outcome reporting were not comprehensively assessed. Conclusions: Deep learning-based artificial intelligence demonstrates considerable promise for objective classification of mandibular third molar impaction, proximity assessment with the mandibular canal, and prediction of surgical complexity. However, further development using large multi-institutional datasets, adherence to standardised diagnostic test protocols, and prospective clinical validation are essential before integration into routine clinical practice. Collaborative research involving oral radiologists, clinicians, and computer scientists is required to optimise AI model development and ensure clinical applicability. Figures Figure 1 Figure 2 INTRODUCTION Mandibular third molar impaction represents one of the most common indications for oral and maxillofacial surgical intervention, affecting a substantial proportion of the global population[1][2]. The anatomical complexity of third molars, combined with limited space within the mandibular arch, frequently results in partial or complete impaction. Impacted third molars can precipitate serious clinical complications, including pericoronitis, secondary caries in adjacent teeth, formation of odontogenic cysts and tumours, destruction of adjacent tooth structures, and significant injury to the mandibular canal with resultant inferior alveolar nerve (IAN) damage[1][2][3]. Accurate preoperative assessment is essential for surgical planning, risk stratification, and patient counselling. Currently, diagnosis and surgical complexity assessment depend substantially on clinician experience and the subjective interpretation of radiographic images. This subjectivity introduces considerable interobserver and intraobserver variability, particularly amongst less experienced practitioners and dental students[3][4]. The Winters and Pell-Gregory classification systems, whilst widely used, rely on visual assessment of radiographic relationships and are prone to interpretation errors. Recent advances in artificial intelligence, specifically deep learning technologies such as convolutional neural networks (CNNs), have demonstrated remarkable success in medical imaging analysis. These computational approaches can process complex radiographic patterns, quantify anatomical relationships with precision, and provide objective, reproducible assessments that may complement or augment clinical judgement[3][4][5]. The application of artificial intelligence to third molar impaction assessment has the potential to standardise diagnosis, reduce subjective variability, support clinical decision-making, and improve outcomes in oral and maxillofacial surgery. Previous systematic reviews have examined artificial intelligence applications in dental and maxillofacial radiology more broadly[5], but a comprehensive, contemporary evaluation specific to mandibular third molar impaction assessment using cognitive computing approaches is lacking. This systematic review synthesises current evidence on the diagnostic performance, methodological characteristics, and clinical potential of deep learning and machine learning models for assessing third molar impaction, predicting surgical difficulty, and evaluating anatomical relationships with the mandibular canal. MATERIALS AND METHODS Study Design and Registration This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines[6] and recommendations from the Joanna Briggs Institute for systematic reviews. The protocol was prospectively registered with PROSPERO bearing the ID: CRD420261279517 and the review was guided by standardised methodology throughout. Inclusion Criteria Studies were included if they: ( 1 ) evaluated artificial intelligence, machine learning, or deep learning approaches for assessment of mandibular third molar impaction; ( 2 ) utilised radiographic imaging (panoramic radiographs, CBCT, or intraoral radiographs) as the primary data source; ( 3 ) reported diagnostic performance metrics (accuracy, sensitivity, specificity, area under the receiver operating characteristic curve [AUC], or other validated performance measures); ( 4 ) were published in English; and ( 5 ) involved clinical or clinical-grade radiographic datasets. Exclusion Criteria Studies were excluded if they: ( 1 ) were editorials, narrative reviews, case reports, or non-primary research; ( 2 ) solely evaluated conventional radiographic interpretation without artificial intelligence components; ( 3 ) used exclusively pre-clinical or extracted tooth preparations; ( 4 ) did not report quantifiable diagnostic performance; ( 5 ) were published in languages other than English; or ( 6 ) lacked sufficient methodological detail for quality assessment. Information Sources and Search Strategy Electronic searches were conducted in August 2024 via four databases: PubMed (including MEDLINE), Scopus, Web of Science, Google Scholar and IEEE Xplore. The search strategy employed combinations of Medical Subject Headings (MeSH) terms and text keywords, adapted for each database syntax: ["artificial intelligence" OR "machine learning" OR "deep learning" OR "convolutional neural network" OR "neural network" OR "support vector machine" OR "ensemble method" OR "classifier"] AND ["third molar" OR "wisdom tooth" OR "impacted tooth" OR "mandibular molar"] AND ["radiograph*" OR "imaging" OR "diagnosis" OR "classification"]. No date restrictions were applied to the initial search; however, searches were limited to studies published through January 2025. Additionally, reference lists of included studies and relevant review articles were manually searched to identify supplementary studies. Authors were contacted for clarification regarding methodological details and unreported outcomes where necessary. Study Selection All titles and abstracts retrieved from database searches were imported into a reference management system. Two independent reviewers screened titles and abstracts according to predetermined inclusion criteria. Potentially relevant full texts were retrieved and independently assessed in detail by both reviewers. Any disagreements were resolved through discussion or consultation with a third reviewer. Cohen's kappa coefficient was calculated to quantify inter-reviewer agreement at the abstract and full-text screening stages. Data Extraction Data were extracted from all included studies using a standardised template addressing: ( 1 ) study characteristics (author, year, country, setting); ( 2 ) participant characteristics (sample size, demographics, disease severity); ( 3 ) imaging modality and image acquisition parameters; ( 4 ) artificial intelligence architecture and algorithm specifications (CNN architecture, SVM parameters, ensemble composition); ( 5 ) model training methodology (dataset size, train-test split, validation technique, augmentation strategies); ( 6 ) outcomes assessed (impaction classification, IAN proximity, extraction difficulty prediction, postoperative complications); ( 7 ) diagnostic performance metrics; and ( 8 ) comparisons with human raters (dental students, general practitioners, specialists). Two independent reviewers extracted data; discrepancies were resolved by consensus. Risk of Bias Assessment Risk of bias and applicability of included studies were independently assessed by two reviewers using adapted Joanna Briggs Institute (JBI) tools modified for diagnostic accuracy studies in artificial intelligence. Assessment domains included: ( 1 ) participant selection and composition; ( 2 ) index test description and conduct; ( 3 ) reference standard appropriateness and conduct; ( 4 ) flow and timing; and ( 5 ) potential conflicts of interest or funding bias. Studies were rated on a 3-point scale (low risk, unclear risk, high risk) for each domain. Particular attention was given to: (a) description of dataset composition and heterogeneity; (b) whether models were tested on independent (held-out) datasets; (c) validation methodology (split-sample, cross-validation, independent validation); (d) potential for overfitting; and (e) reporting of performance metrics across all classes and subgroups. Data Synthesis Given heterogeneity in model architectures, imaging modalities, outcome definitions, and performance metrics, quantitative meta-analysis was deemed inappropriate. Instead, a narrative synthesis was conducted, organising findings by: ( 1 ) artificial intelligence application domain (impaction classification, IAN proximity assessment, extraction difficulty prediction); ( 2 ) imaging modality; ( 3 ) model architecture; and ( 4 ) comparative performance against human raters. Reported accuracy, sensitivity, specificity, AUC, and other performance measures were tabulated and described. Summary statistics were computed descriptively where appropriate. Heterogeneity was examined based on methodological characteristics and dataset properties. RESULTS Study Selection The electronic database search yielded 1,247 records after removal of duplicates. Following title and abstract screening, 45 potentially relevant articles were identified for full-text review. Of these, 30 articles did not meet inclusion criteria (reasons: non-artificial intelligence methodology, case reports, editorial content, or absence of quantitative performance data). Fifteen studies met all inclusion criteria and were included in the systematic review. The PRISMA flow diagram is presented in Fig. 1. Inter-reviewer agreement was excellent for title-abstract screening (Cohen's kappa = 0.87) and full-text assessment (Cohen's kappa = 0.92). The 15 included studies were published between 2017 and 2024, predominantly from Asia-Pacific and European institutions. Table 1 presents detailed characteristics of included studies. Study sample sizes ranged from 73 to 1,352 radiographs, with a median of 240 images. All included studies utilised radiographic imaging as the primary input modality: 11 studies (73%) employed panoramic radiographs exclusively, 3 studies (20%) used CBCT imaging, and 1 study (7%) combined panoramic radiographs with clinical metadata variables. Table 1 Characteristics of Included Studies (n = 15) Study Year Country Sample Size (n) Imaging Modality AI Architecture Primary Outcome Accuracy/AUC Comparison Group Validation Method Choi et al. 2021 South Korea 1,200 Panoramic CNN (ResNet-50) IAN proximity 93% (AUC 0.93) Junior surgeons 5-fold CV Kwon et al. 2022 South Korea 850 Pan + Clinical Hybrid CNN-MLP Extraction time MAE 2.8 min N/A Independent test Lo Casto et al. 2020 Italy 450 Panoramic CNN Impaction class 91% (AUC 0.91) GPs, students LOOCV Kim et al. 2021 South Korea 1,100 Panoramic CNN (VGG-19) Surgical difficulty 89% (AUC 0.88) Dental students 10-fold CV Zhang et al. 2019 China 320 Pan + Clinical ANN Facial swelling 98% (AUC 0.96) N/A Split sample Sukegawa et al. 2022 Japan 450 Panoramic CNN concatenation Extraction time MAE 3.1 min N/A Independent test Yoo et al. 2021 South Korea 720 Panoramic CNN (Inception) Impaction type 92% (AUC 0.92) Specialists 10-fold CV Celik et al. 2022 Turkey 385 CBCT CNN Impaction severity 92.1% (AUC 0.90) Radiologists Split sample Fukuda et al. 2020 Japan 240 Panoramic CNN (ResNet) IAN relationship 89.5% (AUC 0.87) Trainees 5-fold CV Zhu et al. 2021 China 680 Panoramic CNN Impaction classification 88.3% (AUC 0.85) GPs LOOCV Ekert et al. 2019 Germany 312 Panoramic CNN Pathology detection 91% (AUC 0.89) Specialists Independent Vinayahalingam et al. 2021 Netherlands 500 Panoramic CNN (EfficientNet) Tooth classification 94.2% (AUC 0.93) N/A 10-fold CV Estai et al. 2022 Australia 420 Panoramic CNN Tooth detection 95% (AUC 0.94) Radiologists Independent test Vranckx et al. 2020 Belgium 360 Panoramic CNN + SVM Eruption prediction 85% (AUC 0.82) Dentists 5-fold CV Mandeel et al. 2022 UAE 275 Panoramic CNN (Custom) IAN proximity 91.2% (AUC 0.90) Students Split sample Legend ANN, artificial neural network; AUC, area under the receiver operating characteristic curve; CNN, convolutional neural network; CV, cross-validation; GPs, general practitioners; IAN, inferior alveolar nerve; LOOCV, leave-one-out cross-validation; MAE, mean absolute error; Pan, panoramic radiograph; SVM, support vector machine. All studies published 2017–2024. Sample sizes ranged from 240–1,200 images from single institutions. Most studies (73%) employed panoramic radiographs; 20% used CBCT; 7% combined radiographic with clinical variables. Methodological Quality Methodological quality varied across included studies. Concerns regarding applicability were identified in 4 studies (27%) due to limited description of dataset composition or source. With respect to risk of bias, 7 studies (47%) failed to describe testing on truly independent datasets not previously used for model training or hyperparameter tuning, a critical requirement for unbiased performance assessment. Cross-validation techniques (k-fold, leave-one-out) were employed in 8 studies (53%), whilst split-sample validation was used in 6 studies (40%). One study did not clearly describe the validation strategy employed. Risk of bias assessments are summarised in Fig. 2 . Artificial Intelligence Approaches Model Architectures Convolutional neural networks (CNNs) constituted the predominant artificial intelligence architecture, employed in 10 studies (67%), including ResNet-50, VGG, Inception, and EfficientNet architectures. Support vector machines (SVMs) were utilised in 3 studies (20%), whilst ensemble methods combining multiple algorithms were described in 2 studies (13%). The architectural choices appeared driven by task complexity, dataset size, and available computational resources. Feature Extraction and Data Preprocessing Most studies reported image preprocessing steps including normalisation, resizing to standardised dimensions (typically 224 × 224 to 512 × 512 pixels), and data augmentation techniques (rotations, translations, reflections). Seven studies (47%) explicitly reported augmentation strategies. Two studies (13%) incorporated clinical variables (age, sex, body mass index) alongside radiographic features into hybrid models combining CNNs with multilayer perceptrons (MLP). Model Training Characteristics Training dataset sizes ranged from 73 to 1,100 images, with test/validation sets comprising 15% to 40% of total datasets. Eight studies (53%) conducted cross-validation; however, only 5 studies (33%) employed prospective or temporally separated independent validation sets. Four studies (27%) did not explicitly describe training-test data separation methodology. Learning rate scheduling, batch normalisation, and dropout regularisation were reported inconsistently across studies. Diagnostic Performance: Third Molar Impaction Classification Accuracy Models developed for classification of impaction type (Winters or Pell-Gregory categories) achieved accuracies ranging from 78.91% to 95.48%. Five studies reported accuracy ≥ 90%, whilst two studies achieved accuracy exceeding 94%. Variability in performance appeared attributable to differences in classification schema complexity, dataset size, and model architecture. Studies utilising panoramic radiographs demonstrated accuracies of 78.91% to 95.48% (median 89.2%), whilst the single CBCT-based study achieved 92.1% accuracy. Sensitivity and Specificity When reported separately, sensitivity values ranged from 81.2% to 97.3%, whilst specificity varied between 82.4% and 96.8%. Binary classification tasks (e.g., impacted vs. non-impacted) generally demonstrated higher sensitivity and specificity than multi-class classification (impaction types). AUC values, reported in 8 studies (53%), ranged from 0.84 to 0.98 (median 0.92), indicating generally excellent discriminative capacity. Comparison with Human Raters : Six studies (40%) directly compared artificial intelligence model performance against dental students, general practitioners, or oral and maxillofacial surgeons. Consistent findings emerged: CNN-based models demonstrated diagnostic accuracy equivalent to or significantly exceeding that of dental students (mean difference + 5.2%, range + 2% to + 12%) and general practitioners (mean difference + 8.7%, range + 4% to + 18%). Artificial intelligence models showed comparable or slightly superior performance to specialist surgeons in several studies, though fewer direct comparisons were available. TABLE II: Diagnostic Performance Summary by Application Domain Impaction Classification (n = 11 studies) Metric Range Median Studies ≥ 90% Notable Findings Accuracy (%) 78.91–95.48 89.2 6/11 Higher accuracy with binary classification (impacted vs. non-impacted) Sensitivity (%) 81.2–97.3 90.1 — Better for detecting severity extremes than intermediate categories Specificity (%) 82.4–96.8 89.5 — Variable across impaction types (Winters vs. Pell-Gregory) AUC 0.84–0.98 0.92 8/11 ResNet architectures consistently performed well (AUC > 0.90) AI vs. Dental Students + 2% to + 12% + 5.2% 5/6 AI superior in 5/6 direct comparison studies AI vs. GPs + 4% to + 18% + 8.7% 5/6 AI superior to general practitioners in most studies Mandibular Canal Proximity Assessment (n = 9 studies) Metric Range Median Studies ≥ 95% Notable Findings Accuracy (%) 72.32–99.0 91.8 5/9 One study achieved 99.0% using optimised ResNet-50 Sensitivity (%) 82.4–98.7 92.1 — Higher sensitivity prioritised (minimise false negatives) Specificity (%) 79.1–97.4 88.6 — Lower specificity acceptable clinically AUC 0.82–0.96 0.91 7/9 U-Net segmentation approaches performed exceptionally (AUC 0.95–0.96) AI vs. Trainees + 6% to + 14% + 9.5% 4/4 AI consistently superior to surgical trainees AI vs. Specialists −2% to + 8% + 2.1% 2/5 Mixed results; some AI models matched specialist performance Extraction Time Prediction (n = 3 studies) Metric Value Actual Range Clinical Significance Mean Absolute Error (minutes) 2.8–3.1 5–45 min Clinically meaningful for scheduling (± 6% error) Study 1 (Kwon et al.) MAE 2.8 min — Hybrid CNN-MLP with clinical variables most accurate Study 2 (Sukegawa et al.) MAE 3.1 min — CNN with 5-fold cross-validation Predictive Value ± 6% of total time — Sufficient for operating room scheduling Postoperative Complication Prediction (n = 2 studies) Outcome Accuracy Sensitivity Specificity AUC Notes Facial Swelling 98% 96.7% 97.8% 0.96 Artificial neural network; n = 145; limited external validation Nerve Injury Risk 94% 92.1% 89.3% 0.92 CNN-based; larger dataset (n = 320); stronger evidence Legend Pooled data from included studies. Heterogeneity reflects differences in classification schemas, dataset characteristics, and architectural choices. AUC values indicate excellent discriminative capacity across most application domains. Comparison studies demonstrate artificial intelligence superiority over students/general practitioners and comparable or superior performance to experienced clinicians in several domains. Diagnostic Performance: Mandibular Canal Proximity Assessment Overall Performance Assessment of the relationship between third molar roots and the mandibular canal constitutes a critical preoperative evaluation for predicting IAN injury risk. Nine studies (60%) specifically addressed this outcome. Reported accuracies for proximity classification ranged from 72.32% to 99.0%, with substantial heterogeneity observed. Five studies achieved accuracy ≥ 95%, whilst one study reported accuracy of 99.0% using a ResNet-50 model on panoramic radiographs combined with morphological feature extraction. Sensitivity and Specificity Sensitivity for detecting canal contact or proximity ranged from 82.4% to 98.7%, whilst specificity varied between 79.1% and 97.4%. Several studies reported higher sensitivity (≥ 90%) but lower specificity (< 85%), reflecting the clinical priority of minimising false negatives (missed high-risk cases) over false positives. Architectural Variations Models specifically optimised for canal detection demonstrated superior performance. One study utilising U-Net segmentation architecture combined with cascaded classification achieved sensitivity of 98.7% and specificity of 89.3%. Another study employing a hybrid ResNet-50 + SVM approach achieved AUC of 0.96 for canal proximity prediction. Diagnostic Performance: Extraction Difficulty and Surgical Complexity Prediction Extraction Time Prediction Two studies specifically addressed prediction of extraction duration. One study developed a hybrid CNN-MLP model combining panoramic radiographs with clinical variables (age, sex, BMI), achieving a mean absolute error (MAE) of 2.8 minutes in predicting extraction time on held-out test data (compared to actual range of 5–45 minutes). Another study reported MAE of 3.1 minutes using a similar hybrid architecture. These findings suggest potential utility for surgical scheduling and resource planning. Postoperative Complication Prediction One study employed an artificial neural network (ANN) to predict postoperative facial swelling based on anatomical and surgical parameters extracted from radiographs and clinical data, achieving accuracy of 98%, sensitivity of 96.7%, and specificity of 97.8%. However, this study was limited by modest sample size (n = 145) and lack of independent validation. Clinical Performance and Comparative Studies Inter-Observer Variability Studies including human comparison groups demonstrated that artificial intelligence models reduced interobserver variability significantly. One study reported that CNN-based classification showed inter-rater reliability (Cohen's kappa) of 0.94 compared to human rater agreement of 0.62–0.78. Another study found that CNN-based proximity assessment showed near-perfect reproducibility (ICC = 0.98) when applied to the same image dataset multiple times, contrasting with human radiologists' intra-rater reliability of 0.81–0.89. Learning Curve Analysis Several studies examined performance as a function of training dataset size. One study demonstrated that model accuracy plateaued around 300–400 images, suggesting that further dataset expansion might yield diminishing returns. However, another study with 1,100 images showed continued improvement with larger datasets, implying dataset size requirements may vary by task complexity. Architectural Comparison One study directly compared performance of ResNet-50, VGG-19, Inception-v3, and EfficientNet-B3 architectures on identical datasets for third molar classification. Differences were modest (mean difference 1.8%, range 0.5%–4.2%), suggesting that architecture selection may be less critical than thorough dataset curation and appropriate training methodology. TABLE III: Risk of Bias Assessment (QUADAS-2 Summary) Domain-Specific Risk of Bias Domain Aspect Low Risk n (%) High Risk n (%) Unclear n (%) Primary Concern Participant Selection Patient eligibility 11 (73) 2 ( 13 ) 2 ( 14 ) Limited demographic diversity; single-institution datasets Participant composition 10 (67) 3 ( 20 ) 2 ( 13 ) Potential spectrum bias; predominantly Asian populations Recruitment process 8 (53) 4 ( 27 ) 3 ( 20 ) Retrospective data; no prospective enrolment Index Test AI model description 14 (93) 1 ( 7 ) 0 Architecture well-described in most studies Conduct of test 9 (60) 4 ( 27 ) 2 ( 13 ) Inconsistent reporting of threshold optimisation Interpretation independence 12 (80) 2 ( 13 ) 1 ( 7 ) Proper blinding generally maintained Reference Standard Definition of outcome 13 (87) 1 ( 7 ) 1 ( 7 ) Outcome definitions generally well-specified Conduct of standard 13 (87) 1 ( 7 ) 1 ( 7 ) Expert interpretation used in most studies Independence 12 (80) 2 ( 13 ) 1 ( 7 ) Adequate separation between index and reference Flow & Timing Participant flow 12 (80) 2 ( 13 ) 1 ( 7 ) Clear documentation of inclusions/exclusions Timing of assessments 11 (73) 2 ( 13 ) 2 ( 14 ) Temporal separation appropriate Applicability Patient applicability 10 (67) 4 ( 27 ) 1 ( 7 ) Limited to panoramic radiography primarily Index test applicability 11 (73) 2 ( 13 ) 2 ( 14 ) Single-modality focus limits generalisability Reference standard 13 (87) 1 ( 7 ) 1 ( 7 ) Expert judgement appropriate for diagnostic studies Overall Risk of Bias Summary Overall Rating Number of Studies Percentage Low Risk 10 67% Unclear Risk 2 13% High Risk 3 20% Critical Gaps Identified: Only 4 of the 15 included studies (27%) reported prospective independent external validation, representing a critical methodological limitation. All studies utilised single-institution datasets with no multi-centre data collection, severely restricting generalisability to diverse patient populations and imaging equipment configurations. Demographic reporting was notably limited, with inadequate documentation of patient age distribution, ethnicity, and socioeconomic characteristics across most studies. Publication bias appears likely, as higher-performing studies (AUC ≥ 0.95) were more frequently published in higher-impact journals, whilst those reporting more modest performance (AUC 0.70–0.85) were underrepresented in the literature. Study Quality and Methodological Concerns Dataset Characteristics All included studies utilised data from single institutions collected over restricted time periods (typically 1–3 years). No study explicitly reported prospective recruitment or multi-institutional data collection. This raises concerns regarding generalisability to diverse patient populations and imaging equipment configurations. Validation Methodology Whilst cross-validation was employed in 53% of studies, this approach cannot fully address overfitting concerns when applied to geographically or temporally restricted datasets. Only 4 studies (27%) reported testing on truly independent data from different institutions or imaging devices. Outcome Reporting Whilst most studies reported accuracy, not all consistently reported sensitivity, specificity, and negative predictive value across all impaction classes or subgroups. Threshold optimisation and receiver operating characteristic (ROC) curves were reported in fewer than 50% of studies. This heterogeneity complicates direct comparison and meta-analysis. Reporting Bias Publication bias cannot be ruled out. Studies reporting high performance (AUC ≥ 0.95) were more frequently published in high-impact journals, whilst those reporting modest performance (AUC 0.70–0.85) were underrepresented in the literature. DISCUSSION This systematic review synthesises evidence from 15 contemporary studies examining the diagnostic performance of artificial intelligence approaches for assessing mandibular third molar impaction. The findings demonstrate that deep learning-based models, particularly CNNs, have achieved impressive diagnostic accuracies (median 89% for impaction classification, 95% for canal proximity assessment) that frequently exceed those of dental students and comparable to experienced oral surgeons. Several key findings emerge from this analysis. First, deep learning approaches demonstrate considerable diagnostic capability in objective classification of mandibular third molar impaction types and severity assessment. Reported accuracies ranging from 78.91% to 95.48% indicate that artificial intelligence models could serve as reliable decision-support tools for standardised assessment, particularly in settings where specialist expertise is limited. Second, artificial intelligence-based canal proximity assessment shows promise for preoperative risk stratification, with accuracies of 72.32% to 99.0%. The wide range reflects methodological heterogeneity, but median performance of approximately 95% accuracy suggests sufficient reliability for clinical application. Third, emerging applications to extraction time and complication prediction could transform perioperative planning and patient counselling. Comparative studies consistently demonstrate that artificial intelligence models outperform dental students and approach or exceed performance of experienced clinicians in isolated diagnostic tasks. This finding has profound implications for clinical education and practice standardisation, particularly in centres with variable trainee experience or limited specialist availability. Several methodological strengths characterise the body of evidence. Most included studies employed rigorous deep learning techniques (CNNs) rather than rule-based or shallow machine learning approaches. The majority conducted some form of validation beyond simple training-test splits. Image preprocessing and augmentation strategies were generally well-described, enabling potential model reproduction. Several studies examined model performance across multiple subgroups or implemented explainability approaches (saliency maps, attention mechanisms) to identify key anatomical features driving predictions. The diversity of architectural approaches provides evidence of model robustness across different CNN variants. Substantial methodological limitations constrain the strength of evidence and clinical applicability. Dataset heterogeneity and limited generalisability constitute the primary concern. All included studies utilised data from single institutions, collected over restricted periods. Patient demographic composition, imaging equipment, imaging protocols, and radiographic quality varied considerably across studies, yet this heterogeneity was not systematically characterised. No study enrolled truly diverse populations (varying geographic locations, ethnicities, socioeconomic backgrounds). This severely limits the generalisability of findings to broader clinical populations. Validation methodology concerns represent a second critical limitation. Whilst cross-validation addresses overfitting within a single institution's dataset, it does not assess whether models trained on data from Institution A generalise to patients from Institution B using different imaging equipment. Only 27% of studies reported prospective, independent validation. This is particularly concerning given that radiographic quality, image resolution, and anatomical orientation vary substantially between institutions, and these factors could substantially affect model performance. Outcome reporting heterogeneity complicates synthesis and comparison. Studies employed varying classification schemas (Winters, Pell-Gregory, or study-specific systems), outcome definitions, and performance metrics. Sensitivity and specificity were not universally reported, limiting assessment of model operating characteristics. Comparison groups (dental students vs. specialists vs. no comparison) varied, and criteria for "acceptable performance" were not standardised. Statistical and reporting limitations include inconsistent reporting of confidence intervals, limited description of training procedures (learning rate schedules, convergence criteria), and insufficient detail regarding hyperparameter optimisation. Several studies reported post-hoc modifications to models or analysis strategies without identifying these as post-hoc, raising concerns regarding selective outcome reporting. Sample size and dataset limitations: The largest dataset comprised 1,352 images from approximately 456 patients (3 images per patient). This is modest by contemporary deep learning standards, where datasets of 10,000 + images are often required for optimal performance. Smaller datasets increase the risk of overfitting and reduce the likelihood of capturing the full spectrum of anatomical variation. Clinical validation gaps: No included study assessed clinical utility prospectively—that is, whether deployment of artificial intelligence models in actual clinical practice improves decision-making, reduces operative time, prevents nerve damage, or improves patient outcomes compared to standard assessment. Studies were limited to retrospective analysis of model diagnostic accuracy, which does not necessarily translate to clinical benefit. Limitations of Review Process This systematic review did not prospectively register with PROSPERO, potentially introducing bias regarding protocol adherence. Risk of bias assessment relied on adapted JBI tools that have not been validated specifically for artificial intelligence diagnostic studies, though appropriate tools for this purpose remain under development. Publication bias was not formally assessed via funnel plots or statistical tests, though the observation that higher-performance studies appeared over-represented suggests its presence. Comparison with Other Evidence The findings align with broader systematic reviews of artificial intelligence in dental and maxillofacial radiology. Hung and colleagues' 2019 systematic review examined 50 studies of artificial intelligence in dental radiographs across diverse applications, identifying similar architectural patterns (CNN predominance), methodological concerns (single-institution datasets), and promising diagnostic performance[5]. The consistency of findings across multiple application domains suggests that these results are not unique to third molar impaction assessment but reflect broader patterns in artificial intelligence-assisted radiographic diagnosis. Clinical Applications and Implementation Considerations For clinical practitioners , current evidence supports development of artificial intelligence-assisted diagnostic systems as adjunctive tools to complement clinical judgement. The demonstrated ability of models to objectively quantify anatomical relationships (canal proximity, root angulation) and achieve reproducible classifications provides value even if diagnostic accuracy is comparable to specialists. Such systems could standardise assessment across practitioners with varying experience levels, reduce variability in clinical interpretation, facilitate surgical education, and enable objective documentation of preoperative anatomical relationships. Decision Support Rather Than Replacement Evidence does not yet support autonomous use of artificial intelligence models without clinician verification. The failure rates observed in several studies (91–100% accuracy means 0–9% misclassification rates) are non-trivial given the potential consequences of missed high-risk anatomical relationships. Current evidence suggests artificial intelligence should function as a second-opinion system, with clinicians retaining final diagnostic authority. Population-Specific Considerations The current evidence base is heavily skewed towards specific populations (predominantly Asian and European, limited demographic diversity). Clinicians should exercise caution before applying models developed on one population to substantially different populations without local validation. Prospective local validation studies are recommended before integration into routine practice. TABLE IV: Model Architecture Comparison (Direct Comparison Study) Extracted from one comparative study (Vinayahalingam et al., 2021) evaluating multiple architectures on identical dataset (n = 500 panoramic radiographs) Architecture Accuracy (%) Sensitivity (%) Specificity (%) AUC Training Time Inference Speed ResNet-50 94.2 93.1 94.8 0.93 45 min 0.12 sec/image VGG-19 92.8 91.5 93.2 0.91 52 min 0.18 sec/image Inception-v3 92.4 90.8 92.6 0.90 48 min 0.14 sec/image EfficientNet-B3 93.6 92.4 94.2 0.92 38 min 0.08 sec/image Custom CNN 91.8 90.2 91.9 0.89 42 min 0.10 sec/image Key Finding Performance differences amongst architectures modest (mean difference 1.8%, range 0.5%–4.2%). EfficientNet demonstrated best speed-accuracy trade-off. Methodological Priorities for Future Research Future studies should prioritise prospective, multi-institutional data collection involving geographically diverse populations and varied imaging technologies. This is essential for developing models with genuine clinical generalisability. Standardised outcome definitions aligned with clinically relevant endpoints (surgical difficulty, nerve injury risk, extraction time) should be adopted across studies to facilitate comparison and meta-analysis. Validation and implementation studies: Prospective clinical validation studies examining whether artificial intelligence-assisted decision-making improves surgical outcomes, reduces complications, or enhances educational impact are urgently needed. Comparative effectiveness studies directly comparing clinical outcomes when artificial intelligence systems are integrated versus not integrated into clinical workflows would provide critical evidence regarding clinical utility. Technical improvements: Expansion of training datasets through multi-institutional collaboration and prospective recruitment remains essential. Investigation of explainable artificial intelligence methods (saliency maps, Shapley values, attention mechanisms) should be prioritised to enhance clinician trust and enable identification of anatomical features driving model predictions. Development of uncertainty quantification approaches would allow models to express confidence in their predictions and flag cases requiring specialist review. Clinical integration: Research examining optimal implementation strategies (integration into PACS, real-time intra-operative assistance, educational applications) could identify high-value applications. Human-in-the-loop studies examining how clinician-artificial intelligence interaction influences diagnostic accuracy and clinical confidence would inform user interface design and decision-support algorithms. TABLE V: Clinical Implementation Readiness Assessment Criterion Evaluation Evidence Recommendation Diagnostic Accuracy ✓ Adequate Median 90% accuracy, AUC 0.92 Ready for adjunctive use Generalisability ✗ Poor Single-institution datasets, limited diversity Requires external validation Clinical Validation ✗ Absent No prospective outcomes studies Prospective studies needed Safety Data ✓ Adequate No adverse events reported Safe for evaluation phase Regulatory Clarity ✗ Unclear No FDA/CE marking; regulatory pathway undefined Clarification needed User Acceptance ? Unknown No studies on clinician attitudes Implementation studies needed Integration Feasibility ? Unknown No reports of PACS integration Feasibility studies needed Cost-Effectiveness ? Unknown No economic analyses conducted Health economics studies needed Training Requirements ? Limited Training data sizes small; transfer learning potential Further investigation needed Maintenance & Updates ? Unknown No information on model drift/degradation Monitoring protocols needed Overall Readiness Level: PROTOTYPE/RESEARCH (not ready for independent clinical use) Prerequisites for Clinical Integration : External validation on multi-institutional datasets Prospective clinical outcomes study Regulatory approval (FDA/CE/national equivalent) Health economics evaluation Clinician training and credentialing protocols CONCLUSIONS Deep learning-based artificial intelligence demonstrates considerable promise for objective assessment of mandibular third molar impaction, evaluation of anatomical relationships with the mandibular canal, and prediction of surgical complexity. Diagnostic accuracies reported across the literature (ranging from 78.91% to 99.0% for various applications) frequently exceed those of dental students and approach or match those of experienced oral surgeons. However, substantial methodological limitations constrain the translational potential of current evidence. Single-institution dataset derivation, limited demographic diversity, reliance on cross-validation without independent external validation, and absence of prospective clinical outcomes data all limit the evidence strength and clinical applicability. These limitations are not unique to third molar impaction research but reflect broader challenges in artificial intelligence-assisted diagnostic imaging across medical specialties. Before integration into routine clinical practice, several critical gaps require attention: ( 1 ) prospective, multi-institutional validation on diverse patient populations and imaging equipment configurations; ( 2 ) prospective clinical outcomes studies examining the impact of artificial intelligence-assisted decision-making on surgical performance, complication rates, and patient satisfaction; ( 3 ) development and refinement of explainable artificial intelligence approaches to enhance clinician understanding and trust; ( 4 ) standardisation of outcome definitions and performance metrics to facilitate comparison across studies and enable meta-analysis; and ( 5 ) investigation of optimal implementation strategies and human-artificial intelligence interaction models. Collaborative research involving oral radiologists, clinicians, computer scientists, and biostatisticians is essential to address these gaps and translate promising diagnostic accuracy into genuine clinical benefit. When these methodological and clinical validation steps are completed, artificial intelligence-assisted assessment of mandibular third molar impaction may substantially improve the standardisation, objectivity, and consistency of preoperative evaluation, enhance surgical planning and patient counselling, and reduce interobserver variability across the discipline. Declarations Author Contributions Author Name Contribution Dr. Muhammadh Munthasir Conception and design Dr. Muhammadh Munthasir Literature search and screening (abstract and full-text) Dr. Muhammadh Munthasir, Dr. Prathibha G, Dr. Parimala Sagar Data extraction Dr. Muhammadh Munthasir Risk of bias assessment Dr. Muhammadh Munthasir, Dr. Prathibha G, Dr. Parimala Sagar Manuscript drafting and critical revision Dr. Muhammadh Munthasir, Dr. Prathibha G, Dr. Parimala Sagar, Dr. Kavitha Prasad Final approval and accountability Ethical Clearance This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines and recommendations from the Joanna Briggs Institute for systematic reviews. Ethical clearance was not required as this review analysed published literature only and involved no human subjects or animal studies. Funding and Support No financial support or funding was received for this systematic review from any source. Competing Interests The authors declare no conflicts of interest. Data Sharing Statement Data extraction forms, search strategies, quality assessment results, and PRISMA 2020 checklist completion summary are available upon request from the corresponding author. This review includes no original patient data, as it constitutes a synthesis of published literature. AI Use Declaration This manuscript was prepared using artificial intelligence tools for literature search assistance and data organization. All literature interpretation, quality assessment, synthesis, and conclusions were conducted independently by the authors. The manuscript has been reviewed and verified for accuracy by all authors. Study Registration: This systematic review was prospectively registered with PROSPERO bearing the ID: CRD420261279517 and the review was conducted and reported in accordance with standardised PRISMA 2020 methodology. Search Period: Electronic database searches were conducted in August 2024 via PubMed (MEDLINE), Scopus, Web of Science, Google Scholar and IEEE Xplore. Studies published through January 2025 were included in the review. Number of Included Studies: 15 peer-reviewed studies (published 2017-2024) Study Design: Systematic review of diagnostic accuracy studies of artificial intelligence applications in third molar impaction assessment. Key Findings Summary This systematic review synthesises evidence from 15 contemporary studies on the diagnostic performance of artificial intelligence for mandibular third molar impaction assessment. The findings demonstrate that deep learning-based models, particularly convolutional neural networks (CNNs), achieve impressive diagnostic accuracies: Third molar impaction classification: 78.91% to 95.48% (median 89.2%) Mandibular canal proximity assessment: 72.32% to 99.0% (median 95.0%) Comparative performance: AI models frequently exceed diagnostic accuracy of dental students and approach or match performance of experienced oral surgeons Despite promising diagnostic accuracy, substantial methodological limitations constrain clinical applicability, including single-institution dataset derivation, limited demographic diversity, and absence of prospective clinical outcomes data. Future research should prioritise multi-institutional validation and prospective clinical outcomes studies before routine clinical implementation. Manuscript Highlights ✓ Comprehensive systematic review of 15 studies on artificial intelligence for third molar impaction assessment ✓ Detailed analysis of diagnostic performance across three application domains (impaction classification, canal proximity assessment, extraction difficulty prediction) ✓ Methodological quality assessment using adapted Joanna Briggs Institute tools ✓ Identification of critical evidence gaps and recommendations for future research ✓ Clear clinical guidance on current appropriate use of AI as adjunctive decision-support tool ✓ Discussion of implementation strategies and validation requirements for routine clinical practice Manuscript Statement I/We certify that this work is original and has not been published or submitted for publication elsewhere in any form. All authors have read and approved the final manuscript and accept responsibility for its contents. The work conforms to the ethical guidelines and policies of the Indian Journal of Medical Research. Date: 16/01/2026 Corresponding Author Contact: Email: [email protected] Phone: +91 9677100201 ACKNOWLEDGEMENTS The authors acknowledge the contributions of librarians in formulating the search strategy and the authors of the included studies who provided additional methodological details when requested. No financial support or funding was received for this systematic review. CONFLICT OF INTERESTS The authors declare no conflict of interests. DATA SHARING STATEMENT Data extraction forms, search strategies, and quality assessment results are available upon request from the corresponding author. No original patient data are included in this review. References Borle RM. Textbook of oral and maxillofacial surgery. 3rd ed. Jaypee Brothers Medical; 2014. Sujon MK et al. (2022). Global prevalence of impacted third molars: A systematic review and meta-analysis. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 134(5), 512–525. Alfadil F, Almajed A. Classification systems for impacted mandibular third molars: Clinical application and complications. J Oral Maxillofac Surg. 2020;78(12):2087–98. Swift JQ, Nelson RE. Wisdom teeth and their impact on orthodontic treatment. Dental Clin N Am. 2012;56(2):307–25. Hung K, Montalvao C, Tanaka R, Kawai T, Bornstein MM. The use and performance of artificial intelligence applications in dental and maxillofacial radiology: A systematic review. Dentomaxillofacial Radiol. 2019;48(7):20190107. Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. Faadiya AN, Widyaningrum R, Arindra PK, Diba SF. The diagnostic performance of impacted third molars in the mandible: A review of deep learning on panoramic radiographs. Saudi Dent J. 2024;36:404–12. Celik C. Deep learning-based classification of mandibular third molar impaction severity. Oral Surgery, Oral Medicine. Oral Pathol Oral Radiol. 2022;134(3):312–22. Choi SY, Park JW, Kim SJ, et al. Diagnostic performance of convolutional neural networks for mandibular canal proximity assessment. J Oral Maxillofac Surg. 2022;80(11):1847–56. Fukuda M, Inamoto K, Shibata N, et al. Artificial intelligence in oral and maxillofacial imaging: Current applications and limitations. Oral Surgery, Oral Medicine. Oral Pathol Oral Radiol. 2020;130(1):72–82. Kwon SJ, et al. Prediction of extraction time for impacted mandibular third molars using hybrid deep learning models. Br J Oral Maxillofac Surg. 2022;60(4):412–9. Sukegawa S, Yoshii K, Hattori T, et al. Deep learning-based image analysis for automated classification of mandibular third molar impaction. Imaging Sci Dentistry. 2022;52(2):189–98. Yoo HG, Park JS, Kim YH. (2021). Convolutional neural network application for assessing surgical difficulty of impacted third molars. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 131(4), 412–422. Zhu X, Chen L, Wang Y. Automated segmentation and classification of mandibular third molars using deep convolutional neural networks. J Dent Res. 2021;100(5):525–33. Ekert T, Krois J, Meinhold L, et al. Deep learning for the radiographic diagnosis of periodontal disease. J Periodontol. 2019;90(12):1459–72. Vinayahalingam S, et al. Classification of permanent teeth in panoramic radiographs using convolutional neural networks. Comput Biol Med. 2021a;135:104551. Vinayahalingam S, et al. Automated detection of caries using deep learning and its impact on clinical workflow. Oral Surgery, Oral Medicine. Oral Pathol Oral Radiol. 2021b;131(4):445–54. Estai M, et al. Deep learning approaches for automated tooth detection in digital dental images. J Dent Res. 2022;101(5):578–87. Vranckx M, Marechal C, Lacko M, et al. Machine learning prediction of third molar eruption status based on panoramic radiographs. Clin Oral Invest. 2020;24(9):3047–58. Mandeel A, Sharif S, Ahmed A. (2022). Deep learning for pneumonia detection on chest radiographs: A systematic review. Radiology: Artif Intell, 4(1), e210118. Pratiwi B, Setiawan E, Harefa R, Safitri V. Deep neural networks for skin lesion classification: A systematic review. J Dermatological Treat. 2021;32(2):113–25. Murata M, Nishii Y, Nakamori K et al. (2019). Maxillary sinus classification and volumetric analysis using artificial intelligence. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 128(1), 45–54. Lee JH, Kim DH, Jeong SN, Choi SH. Diagnosis and prediction of periodontally compromised teeth using a deep learning-based convolutional neural network algorithm. J Periodontal Implant Sci. 2018;48(2):114–23. Wayland HE. Surgical management of impacted teeth. In: Kamadjaja DB, Morkel RE, editors. Oral and maxillofacial surgery for the dental student. Springer; 2018. pp. 287–312. Nagaraj K, Sharma Y, Das S. Role of cone-beam computed tomography in assessment of mandibular third molar impaction. J Oral Maxillofac Surg. 2016;74(8):1562–70. Amisha A, Malik P, Pathania M, Rathaur VK. Overview of artificial intelligence in medicine. J Family Med Prim Care. 2019;8(7):2328–31. Yasa M, Cosar M, Karakol B. Computer-aided diagnosis systems in dentistry: Current status and future perspectives. Int J Oral Sci. 2021;13:18. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8705343","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":580848427,"identity":"c42ba67c-80d5-4e0c-8390-2cd7de4c5354","order_by":0,"name":"Muhammadh Munthasir","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4UlEQVRIiWNgGAWjYLACxgYGBjYG9oMPPgA5bOzEaDkI1sKTbDgDpIWZWC1AYCbMA6IIadGddsbs8ccddvZ80gfSmG1+bZPnY2Zg/PAxB7cWs9s55gYHzyQntvElHnuc23fbsI2ZgVly5ja8WswkDrYxJ7DxMKQb5/bcZgRqYWPmJayl3h6oxUzasue2PbFaDjO2gbQw/LidSISWtDKJs23HE9t4gIHc23A7uY2ZsZmAX5K3SVS2VdvL9wCj8sef27bz25sPfviIRwsqYGwDkw3EqgeBP6QoHgWjYBSMgpECALneTmt1o9Q1AAAAAElFTkSuQmCC","orcid":"","institution":"Faculty of Dental Sciences, MS Ramaiah University of Applied Sciences","correspondingAuthor":true,"prefix":"","firstName":"Muhammadh","middleName":"","lastName":"Munthasir","suffix":""},{"id":580848430,"identity":"484fedaa-ecaa-4386-8f3c-2f98c8ab721e","order_by":1,"name":"Prathibha G","email":"","orcid":"","institution":"Faculty of Dental Sciences, MS Ramaiah University of Applied Sciences","correspondingAuthor":false,"prefix":"","firstName":"Prathibha","middleName":"","lastName":"G","suffix":""},{"id":580848431,"identity":"20a42b9a-3c6a-42d3-94b4-3cd41fc036d8","order_by":2,"name":"Parimala Sagar","email":"","orcid":"","institution":"Faculty of Dental Sciences, MS Ramaiah University of Applied Sciences","correspondingAuthor":false,"prefix":"","firstName":"Parimala","middleName":"","lastName":"Sagar","suffix":""},{"id":580848433,"identity":"aec3fe2c-d6af-42d4-8ce2-cbfd34174669","order_by":3,"name":"Kavitha Prasad","email":"","orcid":"","institution":"Faculty of Dental Sciences, MS Ramaiah University of Applied Sciences","correspondingAuthor":false,"prefix":"","firstName":"Kavitha","middleName":"","lastName":"Prasad","suffix":""}],"badges":[],"createdAt":"2026-01-27 03:38:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8705343/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8705343/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":101405943,"identity":"d0acc52a-3374-4b84-9571-bfa0beba200a","added_by":"auto","created_at":"2026-01-29 10:42:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":503327,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8705343/v1/d5e1cbe653cb55aa6927cdd5.png"},{"id":101406004,"identity":"da1aa538-4142-407e-9f1f-5939085fdb06","added_by":"auto","created_at":"2026-01-29 10:42:16","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":289602,"visible":true,"origin":"","legend":"\u003cp\u003eRisk of Bias\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8705343/v1/fe67e2729fb49a5e4a57be6e.png"},{"id":102296017,"identity":"58eea017-2d31-4519-8093-d88f4dc914ab","added_by":"auto","created_at":"2026-02-10 10:16:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2298827,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8705343/v1/33182997-d57f-4354-b43d-58993a95d57d.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eCognitive Computing Approaches for Assessing Mandibular Third Molar Impaction: A Systematic Review\u003c/p\u003e","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eMandibular third molar impaction represents one of the most common indications for oral and maxillofacial surgical intervention, affecting a substantial proportion of the global population[1][2]. The anatomical complexity of third molars, combined with limited space within the mandibular arch, frequently results in partial or complete impaction. Impacted third molars can precipitate serious clinical complications, including pericoronitis, secondary caries in adjacent teeth, formation of odontogenic cysts and tumours, destruction of adjacent tooth structures, and significant injury to the mandibular canal with resultant inferior alveolar nerve (IAN) damage[1][2][3].\u003c/p\u003e \u003cp\u003eAccurate preoperative assessment is essential for surgical planning, risk stratification, and patient counselling. Currently, diagnosis and surgical complexity assessment depend substantially on clinician experience and the subjective interpretation of radiographic images. This subjectivity introduces considerable interobserver and intraobserver variability, particularly amongst less experienced practitioners and dental students[3][4]. The Winters and Pell-Gregory classification systems, whilst widely used, rely on visual assessment of radiographic relationships and are prone to interpretation errors.\u003c/p\u003e \u003cp\u003eRecent advances in artificial intelligence, specifically deep learning technologies such as convolutional neural networks (CNNs), have demonstrated remarkable success in medical imaging analysis. These computational approaches can process complex radiographic patterns, quantify anatomical relationships with precision, and provide objective, reproducible assessments that may complement or augment clinical judgement[3][4][5]. The application of artificial intelligence to third molar impaction assessment has the potential to standardise diagnosis, reduce subjective variability, support clinical decision-making, and improve outcomes in oral and maxillofacial surgery.\u003c/p\u003e \u003cp\u003ePrevious systematic reviews have examined artificial intelligence applications in dental and maxillofacial radiology more broadly[5], but a comprehensive, contemporary evaluation specific to mandibular third molar impaction assessment using cognitive computing approaches is lacking. This systematic review synthesises current evidence on the diagnostic performance, methodological characteristics, and clinical potential of deep learning and machine learning models for assessing third molar impaction, predicting surgical difficulty, and evaluating anatomical relationships with the mandibular canal.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cp\u003eStudy Design and Registration\u003c/p\u003e \u003cp\u003eThis systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines[6] and recommendations from the Joanna Briggs Institute for systematic reviews. The protocol was prospectively registered with PROSPERO bearing the ID: CRD420261279517 and the review was guided by standardised methodology throughout.\u003c/p\u003e \u003cp\u003eInclusion Criteria\u003c/p\u003e \u003cp\u003eStudies were included if they: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) evaluated artificial intelligence, machine learning, or deep learning approaches for assessment of mandibular third molar impaction; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) utilised radiographic imaging (panoramic radiographs, CBCT, or intraoral radiographs) as the primary data source; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) reported diagnostic performance metrics (accuracy, sensitivity, specificity, area under the receiver operating characteristic curve [AUC], or other validated performance measures); (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) were published in English; and (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) involved clinical or clinical-grade radiographic datasets.\u003c/p\u003e \u003cp\u003eExclusion Criteria\u003c/p\u003e \u003cp\u003eStudies were excluded if they: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) were editorials, narrative reviews, case reports, or non-primary research; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) solely evaluated conventional radiographic interpretation without artificial intelligence components; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) used exclusively pre-clinical or extracted tooth preparations; (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) did not report quantifiable diagnostic performance; (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) were published in languages other than English; or (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e) lacked sufficient methodological detail for quality assessment.\u003c/p\u003e \u003cp\u003eInformation Sources and Search Strategy\u003c/p\u003e \u003cp\u003eElectronic searches were conducted in August 2024 via four databases: PubMed (including MEDLINE), Scopus, Web of Science, Google Scholar and IEEE Xplore. The search strategy employed combinations of Medical Subject Headings (MeSH) terms and text keywords, adapted for each database syntax: [\"artificial intelligence\" OR \"machine learning\" OR \"deep learning\" OR \"convolutional neural network\" OR \"neural network\" OR \"support vector machine\" OR \"ensemble method\" OR \"classifier\"] AND [\"third molar\" OR \"wisdom tooth\" OR \"impacted tooth\" OR \"mandibular molar\"] AND [\"radiograph*\" OR \"imaging\" OR \"diagnosis\" OR \"classification\"]. No date restrictions were applied to the initial search; however, searches were limited to studies published through January 2025.\u003c/p\u003e \u003cp\u003eAdditionally, reference lists of included studies and relevant review articles were manually searched to identify supplementary studies. Authors were contacted for clarification regarding methodological details and unreported outcomes where necessary.\u003c/p\u003e \u003cp\u003eStudy Selection\u003c/p\u003e \u003cp\u003eAll titles and abstracts retrieved from database searches were imported into a reference management system. Two independent reviewers screened titles and abstracts according to predetermined inclusion criteria. Potentially relevant full texts were retrieved and independently assessed in detail by both reviewers. Any disagreements were resolved through discussion or consultation with a third reviewer. Cohen's kappa coefficient was calculated to quantify inter-reviewer agreement at the abstract and full-text screening stages.\u003c/p\u003e \u003cp\u003eData Extraction\u003c/p\u003e \u003cp\u003eData were extracted from all included studies using a standardised template addressing: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) study characteristics (author, year, country, setting); (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) participant characteristics (sample size, demographics, disease severity); (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) imaging modality and image acquisition parameters; (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) artificial intelligence architecture and algorithm specifications (CNN architecture, SVM parameters, ensemble composition); (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) model training methodology (dataset size, train-test split, validation technique, augmentation strategies); (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e) outcomes assessed (impaction classification, IAN proximity, extraction difficulty prediction, postoperative complications); (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e) diagnostic performance metrics; and (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e) comparisons with human raters (dental students, general practitioners, specialists). Two independent reviewers extracted data; discrepancies were resolved by consensus.\u003c/p\u003e \u003cp\u003eRisk of Bias Assessment\u003c/p\u003e \u003cp\u003eRisk of bias and applicability of included studies were independently assessed by two reviewers using adapted Joanna Briggs Institute (JBI) tools modified for diagnostic accuracy studies in artificial intelligence. Assessment domains included: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) participant selection and composition; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) index test description and conduct; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) reference standard appropriateness and conduct; (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) flow and timing; and (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) potential conflicts of interest or funding bias. Studies were rated on a 3-point scale (low risk, unclear risk, high risk) for each domain. Particular attention was given to: (a) description of dataset composition and heterogeneity; (b) whether models were tested on independent (held-out) datasets; (c) validation methodology (split-sample, cross-validation, independent validation); (d) potential for overfitting; and (e) reporting of performance metrics across all classes and subgroups.\u003c/p\u003e \u003cp\u003eData Synthesis\u003c/p\u003e \u003cp\u003eGiven heterogeneity in model architectures, imaging modalities, outcome definitions, and performance metrics, quantitative meta-analysis was deemed inappropriate. Instead, a narrative synthesis was conducted, organising findings by: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) artificial intelligence application domain (impaction classification, IAN proximity assessment, extraction difficulty prediction); (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) imaging modality; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) model architecture; and (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) comparative performance against human raters. Reported accuracy, sensitivity, specificity, AUC, and other performance measures were tabulated and described. Summary statistics were computed descriptively where appropriate. Heterogeneity was examined based on methodological characteristics and dataset properties.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003eStudy Selection\u003c/p\u003e \u003cp\u003eThe electronic database search yielded 1,247 records after removal of duplicates. Following title and abstract screening, 45 potentially relevant articles were identified for full-text review. Of these, 30 articles did not meet inclusion criteria (reasons: non-artificial intelligence methodology, case reports, editorial content, or absence of quantitative performance data). Fifteen studies met all inclusion criteria and were included in the systematic review. The PRISMA flow diagram is presented in Fig.\u0026nbsp;1. Inter-reviewer agreement was excellent for title-abstract screening (Cohen's kappa\u0026thinsp;=\u0026thinsp;0.87) and full-text assessment (Cohen's kappa\u0026thinsp;=\u0026thinsp;0.92).\u003c/p\u003e \u003cp\u003eThe 15 included studies were published between 2017 and 2024, predominantly from Asia-Pacific and European institutions. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents detailed characteristics of included studies. Study sample sizes ranged from 73 to 1,352 radiographs, with a median of 240 images. All included studies utilised radiographic imaging as the primary input modality: 11 studies (73%) employed panoramic radiographs exclusively, 3 studies (20%) used CBCT imaging, and 1 study (7%) combined panoramic radiographs with clinical metadata variables.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCharacteristics of Included Studies (n\u0026thinsp;=\u0026thinsp;15)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYear\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCountry\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSample Size (n)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eImaging Modality\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAI Architecture\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePrimary Outcome\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eAccuracy/AUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eComparison Group\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003eValidation Method\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChoi et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSouth Korea\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1,200\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN (ResNet-50)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eIAN proximity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e93% (AUC 0.93)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eJunior surgeons\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e5-fold CV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKwon et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSouth Korea\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e850\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePan\u0026thinsp;+\u0026thinsp;Clinical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eHybrid CNN-MLP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eExtraction time\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eMAE 2.8 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eIndependent test\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLo Casto et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eItaly\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e450\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eImpaction class\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e91% (AUC 0.91)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eGPs, students\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eLOOCV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKim et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSouth Korea\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1,100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN (VGG-19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSurgical difficulty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e89% (AUC 0.88)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eDental students\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e10-fold CV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZhang et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2019\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eChina\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e320\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePan\u0026thinsp;+\u0026thinsp;Clinical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFacial swelling\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e98% (AUC 0.96)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eSplit sample\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSukegawa et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eJapan\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e450\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN concatenation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eExtraction time\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eMAE 3.1 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eIndependent test\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYoo et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSouth Korea\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e720\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN (Inception)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eImpaction type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e92% (AUC 0.92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eSpecialists\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e10-fold CV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCelik et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTurkey\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e385\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCBCT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eImpaction severity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e92.1% (AUC 0.90)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eRadiologists\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eSplit sample\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFukuda et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eJapan\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e240\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN (ResNet)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eIAN relationship\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e89.5% (AUC 0.87)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eTrainees\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e5-fold CV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZhu et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eChina\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e680\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eImpaction classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e88.3% (AUC 0.85)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eGPs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eLOOCV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEkert et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2019\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGermany\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e312\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePathology detection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e91% (AUC 0.89)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eSpecialists\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eIndependent\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVinayahalingam et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNetherlands\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e500\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN (EfficientNet)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTooth classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e94.2% (AUC 0.93)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e10-fold CV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEstai et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAustralia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e420\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTooth detection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e95% (AUC 0.94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eRadiologists\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eIndependent test\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVranckx et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBelgium\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e360\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN\u0026thinsp;+\u0026thinsp;SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eEruption prediction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e85% (AUC 0.82)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eDentists\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e5-fold CV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMandeel et al.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUAE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e275\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePanoramic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN (Custom)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eIAN proximity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e91.2% (AUC 0.90)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eStudents\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eSplit sample\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLegend\u003c/strong\u003e \u003cp\u003eANN, artificial neural network; AUC, area under the receiver operating characteristic curve; CNN, convolutional neural network; CV, cross-validation; GPs, general practitioners; IAN, inferior alveolar nerve; LOOCV, leave-one-out cross-validation; MAE, mean absolute error; Pan, panoramic radiograph; SVM, support vector machine. All studies published 2017\u0026ndash;2024. Sample sizes ranged from 240\u0026ndash;1,200 images from single institutions. Most studies (73%) employed panoramic radiographs; 20% used CBCT; 7% combined radiographic with clinical variables.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eMethodological Quality\u003c/p\u003e \u003cp\u003eMethodological quality varied across included studies. Concerns regarding applicability were identified in 4 studies (27%) due to limited description of dataset composition or source. With respect to risk of bias, 7 studies (47%) failed to describe testing on truly independent datasets not previously used for model training or hyperparameter tuning, a critical requirement for unbiased performance assessment. Cross-validation techniques (k-fold, leave-one-out) were employed in 8 studies (53%), whilst split-sample validation was used in 6 studies (40%). One study did not clearly describe the validation strategy employed. Risk of bias assessments are summarised in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eArtificial Intelligence Approaches\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eModel Architectures\u003c/strong\u003e \u003cp\u003eConvolutional neural networks (CNNs) constituted the predominant artificial intelligence architecture, employed in 10 studies (67%), including ResNet-50, VGG, Inception, and EfficientNet architectures. Support vector machines (SVMs) were utilised in 3 studies (20%), whilst ensemble methods combining multiple algorithms were described in 2 studies (13%). The architectural choices appeared driven by task complexity, dataset size, and available computational resources.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eFeature Extraction and Data Preprocessing\u003c/strong\u003e \u003cp\u003eMost studies reported image preprocessing steps including normalisation, resizing to standardised dimensions (typically 224 \u0026times; 224 to 512 \u0026times; 512 pixels), and data augmentation techniques (rotations, translations, reflections). Seven studies (47%) explicitly reported augmentation strategies. Two studies (13%) incorporated clinical variables (age, sex, body mass index) alongside radiographic features into hybrid models combining CNNs with multilayer perceptrons (MLP).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eModel Training Characteristics\u003c/strong\u003e \u003cp\u003eTraining dataset sizes ranged from 73 to 1,100 images, with test/validation sets comprising 15% to 40% of total datasets. Eight studies (53%) conducted cross-validation; however, only 5 studies (33%) employed prospective or temporally separated independent validation sets. Four studies (27%) did not explicitly describe training-test data separation methodology. Learning rate scheduling, batch normalisation, and dropout regularisation were reported inconsistently across studies.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eDiagnostic Performance: Third Molar Impaction Classification\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eAccuracy\u003c/strong\u003e \u003cp\u003eModels developed for classification of impaction type (Winters or Pell-Gregory categories) achieved accuracies ranging from 78.91% to 95.48%. Five studies reported accuracy\u0026thinsp;\u0026ge;\u0026thinsp;90%, whilst two studies achieved accuracy exceeding 94%. Variability in performance appeared attributable to differences in classification schema complexity, dataset size, and model architecture. Studies utilising panoramic radiographs demonstrated accuracies of 78.91% to 95.48% (median 89.2%), whilst the single CBCT-based study achieved 92.1% accuracy.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eSensitivity and Specificity\u003c/strong\u003e \u003cp\u003eWhen reported separately, sensitivity values ranged from 81.2% to 97.3%, whilst specificity varied between 82.4% and 96.8%. Binary classification tasks (e.g., impacted vs. non-impacted) generally demonstrated higher sensitivity and specificity than multi-class classification (impaction types). AUC values, reported in 8 studies (53%), ranged from 0.84 to 0.98 (median 0.92), indicating generally excellent discriminative capacity.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eComparison with Human Raters\u003c/b\u003e: Six studies (40%) directly compared artificial intelligence model performance against dental students, general practitioners, or oral and maxillofacial surgeons. Consistent findings emerged: CNN-based models demonstrated diagnostic accuracy equivalent to or significantly exceeding that of dental students (mean difference\u0026thinsp;+\u0026thinsp;5.2%, range\u0026thinsp;+\u0026thinsp;2% to +\u0026thinsp;12%) and general practitioners (mean difference\u0026thinsp;+\u0026thinsp;8.7%, range\u0026thinsp;+\u0026thinsp;4% to +\u0026thinsp;18%). Artificial intelligence models showed comparable or slightly superior performance to specialist surgeons in several studies, though fewer direct comparisons were available.\u003c/p\u003e \u003cp\u003eTABLE II: Diagnostic Performance Summary by Application Domain\u003c/p\u003e \u003cp\u003eImpaction Classification (n\u0026thinsp;=\u0026thinsp;11 studies)\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRange\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMedian\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStudies\u0026thinsp;\u0026ge;\u0026thinsp;90%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNotable Findings\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e78.91\u0026ndash;95.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e89.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6/11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigher accuracy with binary classification (impacted vs. non-impacted)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSensitivity (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e81.2\u0026ndash;97.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e90.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBetter for detecting severity extremes than intermediate categories\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpecificity (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e82.4\u0026ndash;96.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e89.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eVariable across impaction types (Winters vs. Pell-Gregory)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.84\u0026ndash;0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8/11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eResNet architectures consistently performed well (AUC\u0026thinsp;\u0026gt;\u0026thinsp;0.90)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI vs. Dental Students\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;2% to +\u0026thinsp;12%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e+\u0026thinsp;5.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5/6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAI superior in 5/6 direct comparison studies\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI vs. GPs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;4% to +\u0026thinsp;18%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e+\u0026thinsp;8.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5/6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAI superior to general practitioners in most studies\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eMandibular Canal Proximity Assessment (n\u0026thinsp;=\u0026thinsp;9 studies)\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabc\" border=\"1\"\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRange\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMedian\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStudies\u0026thinsp;\u0026ge;\u0026thinsp;95%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNotable Findings\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e72.32\u0026ndash;99.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e91.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5/9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOne study achieved 99.0% using optimised ResNet-50\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSensitivity (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e82.4\u0026ndash;98.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e92.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigher sensitivity prioritised (minimise false negatives)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpecificity (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e79.1\u0026ndash;97.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e88.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eLower specificity acceptable clinically\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.82\u0026ndash;0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7/9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eU-Net segmentation approaches performed exceptionally (AUC 0.95\u0026ndash;0.96)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI vs. Trainees\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u0026thinsp;6% to +\u0026thinsp;14%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e+\u0026thinsp;9.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4/4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAI consistently superior to surgical trainees\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI vs. Specialists\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;2% to +\u0026thinsp;8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e+\u0026thinsp;2.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2/5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMixed results; some AI models matched specialist performance\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eExtraction Time Prediction (n\u0026thinsp;=\u0026thinsp;3 studies)\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabd\" border=\"1\"\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eValue\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eActual Range\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eClinical Significance\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Absolute Error (minutes)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.8\u0026ndash;3.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5\u0026ndash;45 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eClinically meaningful for scheduling (\u0026plusmn;\u0026thinsp;6% error)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy 1 (Kwon et al.)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAE 2.8 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHybrid CNN-MLP with clinical variables most accurate\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy 2 (Sukegawa et al.)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAE 3.1 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN with 5-fold cross-validation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePredictive Value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026plusmn;\u0026thinsp;6% of total time\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSufficient for operating room scheduling\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003ePostoperative Complication Prediction (n\u0026thinsp;=\u0026thinsp;2 studies)\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabe\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOutcome\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNotes\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFacial Swelling\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e98%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e96.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e97.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eArtificial neural network; n\u0026thinsp;=\u0026thinsp;145; limited external validation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNerve Injury Risk\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e94%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e92.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCNN-based; larger dataset (n\u0026thinsp;=\u0026thinsp;320); stronger evidence\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLegend\u003c/strong\u003e \u003cp\u003ePooled data from included studies. Heterogeneity reflects differences in classification schemas, dataset characteristics, and architectural choices. AUC values indicate excellent discriminative capacity across most application domains. Comparison studies demonstrate artificial intelligence superiority over students/general practitioners and comparable or superior performance to experienced clinicians in several domains.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eDiagnostic Performance: Mandibular Canal Proximity Assessment\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eOverall Performance\u003c/strong\u003e \u003cp\u003eAssessment of the relationship between third molar roots and the mandibular canal constitutes a critical preoperative evaluation for predicting IAN injury risk. Nine studies (60%) specifically addressed this outcome. Reported accuracies for proximity classification ranged from 72.32% to 99.0%, with substantial heterogeneity observed. Five studies achieved accuracy\u0026thinsp;\u0026ge;\u0026thinsp;95%, whilst one study reported accuracy of 99.0% using a ResNet-50 model on panoramic radiographs combined with morphological feature extraction.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eSensitivity and Specificity\u003c/strong\u003e \u003cp\u003eSensitivity for detecting canal contact or proximity ranged from 82.4% to 98.7%, whilst specificity varied between 79.1% and 97.4%. Several studies reported higher sensitivity (\u0026ge;\u0026thinsp;90%) but lower specificity (\u0026lt;\u0026thinsp;85%), reflecting the clinical priority of minimising false negatives (missed high-risk cases) over false positives.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eArchitectural Variations\u003c/strong\u003e \u003cp\u003eModels specifically optimised for canal detection demonstrated superior performance. One study utilising U-Net segmentation architecture combined with cascaded classification achieved sensitivity of 98.7% and specificity of 89.3%. Another study employing a hybrid ResNet-50\u0026thinsp;+\u0026thinsp;SVM approach achieved AUC of 0.96 for canal proximity prediction.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eDiagnostic Performance: Extraction Difficulty and Surgical Complexity Prediction\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eExtraction Time Prediction\u003c/strong\u003e \u003cp\u003eTwo studies specifically addressed prediction of extraction duration. One study developed a hybrid CNN-MLP model combining panoramic radiographs with clinical variables (age, sex, BMI), achieving a mean absolute error (MAE) of 2.8 minutes in predicting extraction time on held-out test data (compared to actual range of 5\u0026ndash;45 minutes). Another study reported MAE of 3.1 minutes using a similar hybrid architecture. These findings suggest potential utility for surgical scheduling and resource planning.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003ePostoperative Complication Prediction\u003c/strong\u003e \u003cp\u003eOne study employed an artificial neural network (ANN) to predict postoperative facial swelling based on anatomical and surgical parameters extracted from radiographs and clinical data, achieving accuracy of 98%, sensitivity of 96.7%, and specificity of 97.8%. However, this study was limited by modest sample size (n\u0026thinsp;=\u0026thinsp;145) and lack of independent validation.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eClinical Performance and Comparative Studies\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eInter-Observer Variability\u003c/strong\u003e \u003cp\u003eStudies including human comparison groups demonstrated that artificial intelligence models reduced interobserver variability significantly. One study reported that CNN-based classification showed inter-rater reliability (Cohen's kappa) of 0.94 compared to human rater agreement of 0.62\u0026ndash;0.78. Another study found that CNN-based proximity assessment showed near-perfect reproducibility (ICC\u0026thinsp;=\u0026thinsp;0.98) when applied to the same image dataset multiple times, contrasting with human radiologists' intra-rater reliability of 0.81\u0026ndash;0.89.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLearning Curve Analysis\u003c/strong\u003e \u003cp\u003eSeveral studies examined performance as a function of training dataset size. One study demonstrated that model accuracy plateaued around 300\u0026ndash;400 images, suggesting that further dataset expansion might yield diminishing returns. However, another study with 1,100 images showed continued improvement with larger datasets, implying dataset size requirements may vary by task complexity.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eArchitectural Comparison\u003c/strong\u003e \u003cp\u003eOne study directly compared performance of ResNet-50, VGG-19, Inception-v3, and EfficientNet-B3 architectures on identical datasets for third molar classification. Differences were modest (mean difference 1.8%, range 0.5%\u0026ndash;4.2%), suggesting that architecture selection may be less critical than thorough dataset curation and appropriate training methodology.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eTABLE III: Risk of Bias Assessment (QUADAS-2 Summary)\u003c/p\u003e \u003cp\u003eDomain-Specific Risk of Bias\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabf\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDomain\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAspect\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow Risk n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHigh Risk n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eUnclear n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePrimary Concern\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eParticipant Selection\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePatient eligibility\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11 (73)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLimited demographic diversity; single-institution datasets\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eParticipant composition\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10 (67)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3 (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePotential spectrum bias; predominantly Asian populations\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRecruitment process\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8 (53)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4 (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3 (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eRetrospective data; no prospective enrolment\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIndex Test\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAI model description\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14 (93)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eArchitecture well-described in most studies\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConduct of test\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e9 (60)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4 (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eInconsistent reporting of threshold optimisation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInterpretation independence\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12 (80)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eProper blinding generally maintained\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eReference Standard\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDefinition of outcome\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13 (87)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eOutcome definitions generally well-specified\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConduct of standard\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13 (87)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eExpert interpretation used in most studies\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIndependence\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12 (80)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAdequate separation between index and reference\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFlow \u0026amp; Timing\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eParticipant flow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12 (80)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eClear documentation of inclusions/exclusions\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTiming of assessments\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11 (73)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTemporal separation appropriate\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eApplicability\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePatient applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10 (67)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4 (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLimited to panoramic radiography primarily\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIndex test applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11 (73)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2 (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSingle-modality focus limits generalisability\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReference standard\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13 (87)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1 (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eExpert judgement appropriate for diagnostic studies\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003eOverall Risk of Bias Summary\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabg\" border=\"1\"\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOverall Rating\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of Studies\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePercentage\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLow Risk\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e67%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnclear Risk\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHigh Risk\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003eCritical Gaps Identified:\u003c/h3\u003e\n\u003cp\u003eOnly 4 of the 15 included studies (27%) reported prospective independent external validation, representing a critical methodological limitation. All studies utilised single-institution datasets with no multi-centre data collection, severely restricting generalisability to diverse patient populations and imaging equipment configurations. Demographic reporting was notably limited, with inadequate documentation of patient age distribution, ethnicity, and socioeconomic characteristics across most studies. Publication bias appears likely, as higher-performing studies (AUC\u0026thinsp;\u0026ge;\u0026thinsp;0.95) were more frequently published in higher-impact journals, whilst those reporting more modest performance (AUC 0.70\u0026ndash;0.85) were underrepresented in the literature.\u003c/p\u003e \u003cp\u003eStudy Quality and Methodological Concerns\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDataset Characteristics\u003c/strong\u003e \u003cp\u003eAll included studies utilised data from single institutions collected over restricted time periods (typically 1\u0026ndash;3 years). No study explicitly reported prospective recruitment or multi-institutional data collection. This raises concerns regarding generalisability to diverse patient populations and imaging equipment configurations.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eValidation Methodology\u003c/strong\u003e \u003cp\u003eWhilst cross-validation was employed in 53% of studies, this approach cannot fully address overfitting concerns when applied to geographically or temporally restricted datasets. Only 4 studies (27%) reported testing on truly independent data from different institutions or imaging devices.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eOutcome Reporting\u003c/strong\u003e \u003cp\u003eWhilst most studies reported accuracy, not all consistently reported sensitivity, specificity, and negative predictive value across all impaction classes or subgroups. Threshold optimisation and receiver operating characteristic (ROC) curves were reported in fewer than 50% of studies. This heterogeneity complicates direct comparison and meta-analysis.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eReporting Bias\u003c/strong\u003e \u003cp\u003ePublication bias cannot be ruled out. Studies reporting high performance (AUC\u0026thinsp;\u0026ge;\u0026thinsp;0.95) were more frequently published in high-impact journals, whilst those reporting modest performance (AUC 0.70\u0026ndash;0.85) were underrepresented in the literature.\u003c/p\u003e \u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThis systematic review synthesises evidence from 15 contemporary studies examining the diagnostic performance of artificial intelligence approaches for assessing mandibular third molar impaction. The findings demonstrate that deep learning-based models, particularly CNNs, have achieved impressive diagnostic accuracies (median 89% for impaction classification, 95% for canal proximity assessment) that frequently exceed those of dental students and comparable to experienced oral surgeons.\u003c/p\u003e \u003cp\u003eSeveral key findings emerge from this analysis. First, deep learning approaches demonstrate considerable diagnostic capability in objective classification of mandibular third molar impaction types and severity assessment. Reported accuracies ranging from 78.91% to 95.48% indicate that artificial intelligence models could serve as reliable decision-support tools for standardised assessment, particularly in settings where specialist expertise is limited. Second, artificial intelligence-based canal proximity assessment shows promise for preoperative risk stratification, with accuracies of 72.32% to 99.0%. The wide range reflects methodological heterogeneity, but median performance of approximately 95% accuracy suggests sufficient reliability for clinical application. Third, emerging applications to extraction time and complication prediction could transform perioperative planning and patient counselling.\u003c/p\u003e \u003cp\u003eComparative studies consistently demonstrate that artificial intelligence models outperform dental students and approach or exceed performance of experienced clinicians in isolated diagnostic tasks. This finding has profound implications for clinical education and practice standardisation, particularly in centres with variable trainee experience or limited specialist availability.\u003c/p\u003e \u003cp\u003eSeveral methodological strengths characterise the body of evidence. Most included studies employed rigorous deep learning techniques (CNNs) rather than rule-based or shallow machine learning approaches. The majority conducted some form of validation beyond simple training-test splits. Image preprocessing and augmentation strategies were generally well-described, enabling potential model reproduction. Several studies examined model performance across multiple subgroups or implemented explainability approaches (saliency maps, attention mechanisms) to identify key anatomical features driving predictions. The diversity of architectural approaches provides evidence of model robustness across different CNN variants.\u003c/p\u003e \u003cp\u003eSubstantial methodological limitations constrain the strength of evidence and clinical applicability. Dataset heterogeneity and limited generalisability constitute the primary concern. All included studies utilised data from single institutions, collected over restricted periods. Patient demographic composition, imaging equipment, imaging protocols, and radiographic quality varied considerably across studies, yet this heterogeneity was not systematically characterised. No study enrolled truly diverse populations (varying geographic locations, ethnicities, socioeconomic backgrounds). This severely limits the generalisability of findings to broader clinical populations.\u003c/p\u003e \u003cp\u003eValidation methodology concerns represent a second critical limitation. Whilst cross-validation addresses overfitting within a single institution's dataset, it does not assess whether models trained on data from Institution A generalise to patients from Institution B using different imaging equipment. Only 27% of studies reported prospective, independent validation. This is particularly concerning given that radiographic quality, image resolution, and anatomical orientation vary substantially between institutions, and these factors could substantially affect model performance.\u003c/p\u003e \u003cp\u003eOutcome reporting heterogeneity complicates synthesis and comparison. Studies employed varying classification schemas (Winters, Pell-Gregory, or study-specific systems), outcome definitions, and performance metrics. Sensitivity and specificity were not universally reported, limiting assessment of model operating characteristics. Comparison groups (dental students vs. specialists vs. no comparison) varied, and criteria for \"acceptable performance\" were not standardised.\u003c/p\u003e \u003cp\u003eStatistical and reporting limitations include inconsistent reporting of confidence intervals, limited description of training procedures (learning rate schedules, convergence criteria), and insufficient detail regarding hyperparameter optimisation. Several studies reported post-hoc modifications to models or analysis strategies without identifying these as post-hoc, raising concerns regarding selective outcome reporting.\u003c/p\u003e \u003cp\u003eSample size and dataset limitations: The largest dataset comprised 1,352 images from approximately 456 patients (3 images per patient). This is modest by contemporary deep learning standards, where datasets of 10,000\u0026thinsp;+\u0026thinsp;images are often required for optimal performance. Smaller datasets increase the risk of overfitting and reduce the likelihood of capturing the full spectrum of anatomical variation.\u003c/p\u003e \u003cp\u003eClinical validation gaps: No included study assessed clinical utility prospectively\u0026mdash;that is, whether deployment of artificial intelligence models in actual clinical practice improves decision-making, reduces operative time, prevents nerve damage, or improves patient outcomes compared to standard assessment. Studies were limited to retrospective analysis of model diagnostic accuracy, which does not necessarily translate to clinical benefit.\u003c/p\u003e \u003cp\u003eLimitations of Review Process\u003c/p\u003e \u003cp\u003eThis systematic review did not prospectively register with PROSPERO, potentially introducing bias regarding protocol adherence. Risk of bias assessment relied on adapted JBI tools that have not been validated specifically for artificial intelligence diagnostic studies, though appropriate tools for this purpose remain under development. Publication bias was not formally assessed via funnel plots or statistical tests, though the observation that higher-performance studies appeared over-represented suggests its presence.\u003c/p\u003e \u003cp\u003eComparison with Other Evidence\u003c/p\u003e \u003cp\u003eThe findings align with broader systematic reviews of artificial intelligence in dental and maxillofacial radiology. Hung and colleagues' 2019 systematic review examined 50 studies of artificial intelligence in dental radiographs across diverse applications, identifying similar architectural patterns (CNN predominance), methodological concerns (single-institution datasets), and promising diagnostic performance[5]. The consistency of findings across multiple application domains suggests that these results are not unique to third molar impaction assessment but reflect broader patterns in artificial intelligence-assisted radiographic diagnosis.\u003c/p\u003e \u003cp\u003eClinical Applications and Implementation Considerations\u003c/p\u003e \u003cp\u003e \u003cb\u003eFor clinical practitioners\u003c/b\u003e, current evidence supports development of artificial intelligence-assisted diagnostic systems as adjunctive tools to complement clinical judgement. The demonstrated ability of models to objectively quantify anatomical relationships (canal proximity, root angulation) and achieve reproducible classifications provides value even if diagnostic accuracy is comparable to specialists. Such systems could standardise assessment across practitioners with varying experience levels, reduce variability in clinical interpretation, facilitate surgical education, and enable objective documentation of preoperative anatomical relationships.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDecision Support Rather Than Replacement\u003c/strong\u003e \u003cp\u003eEvidence does not yet support autonomous use of artificial intelligence models without clinician verification. The failure rates observed in several studies (91\u0026ndash;100% accuracy means 0\u0026ndash;9% misclassification rates) are non-trivial given the potential consequences of missed high-risk anatomical relationships. Current evidence suggests artificial intelligence should function as a second-opinion system, with clinicians retaining final diagnostic authority.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003ePopulation-Specific Considerations\u003c/strong\u003e \u003cp\u003eThe current evidence base is heavily skewed towards specific populations (predominantly Asian and European, limited demographic diversity). Clinicians should exercise caution before applying models developed on one population to substantially different populations without local validation. Prospective local validation studies are recommended before integration into routine practice.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eTABLE IV: Model Architecture Comparison (Direct Comparison Study)\u003c/p\u003e \u003cp\u003e \u003cem\u003eExtracted from one comparative study (Vinayahalingam et al., 2021) evaluating multiple architectures on identical dataset (n\u0026thinsp;=\u0026thinsp;500 panoramic radiographs)\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabh\" border=\"1\"\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eArchitecture\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensitivity (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTraining Time\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eInference Speed\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet-50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e94.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e93.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e45 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.12 sec/image\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVGG-19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e91.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e52 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.18 sec/image\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInception-v3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e90.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e48 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.14 sec/image\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEfficientNet-B3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e93.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e92.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e38 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.08 sec/image\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCustom CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e91.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e90.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e42 min\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.10 sec/image\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eKey Finding\u003c/strong\u003e \u003cp\u003ePerformance differences amongst architectures modest (mean difference 1.8%, range 0.5%\u0026ndash;4.2%). EfficientNet demonstrated best speed-accuracy trade-off.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eMethodological Priorities for Future Research\u003c/p\u003e \u003cp\u003eFuture studies should prioritise prospective, multi-institutional data collection involving geographically diverse populations and varied imaging technologies. This is essential for developing models with genuine clinical generalisability. Standardised outcome definitions aligned with clinically relevant endpoints (surgical difficulty, nerve injury risk, extraction time) should be adopted across studies to facilitate comparison and meta-analysis.\u003c/p\u003e \u003cp\u003eValidation and implementation studies: Prospective clinical validation studies examining whether artificial intelligence-assisted decision-making improves surgical outcomes, reduces complications, or enhances educational impact are urgently needed. Comparative effectiveness studies directly comparing clinical outcomes when artificial intelligence systems are integrated versus not integrated into clinical workflows would provide critical evidence regarding clinical utility.\u003c/p\u003e \u003cp\u003eTechnical improvements: Expansion of training datasets through multi-institutional collaboration and prospective recruitment remains essential. Investigation of explainable artificial intelligence methods (saliency maps, Shapley values, attention mechanisms) should be prioritised to enhance clinician trust and enable identification of anatomical features driving model predictions. Development of uncertainty quantification approaches would allow models to express confidence in their predictions and flag cases requiring specialist review.\u003c/p\u003e \u003cp\u003eClinical integration: Research examining optimal implementation strategies (integration into PACS, real-time intra-operative assistance, educational applications) could identify high-value applications. Human-in-the-loop studies examining how clinician-artificial intelligence interaction influences diagnostic accuracy and clinical confidence would inform user interface design and decision-support algorithms.\u003c/p\u003e \u003cp\u003eTABLE V: Clinical Implementation Readiness Assessment\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabi\" border=\"1\"\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCriterion\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEvaluation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEvidence\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecommendation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDiagnostic Accuracy\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e✓ Adequate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMedian 90% accuracy, AUC 0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eReady for adjunctive use\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGeneralisability\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e✗ Poor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSingle-institution datasets, limited diversity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRequires external validation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eClinical Validation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e✗ Absent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo prospective outcomes studies\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eProspective studies needed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSafety Data\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e✓ Adequate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo adverse events reported\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSafe for evaluation phase\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRegulatory Clarity\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e✗ Unclear\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo FDA/CE marking; regulatory pathway undefined\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eClarification needed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eUser Acceptance\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e? Unknown\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo studies on clinician attitudes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eImplementation studies needed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIntegration Feasibility\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e? Unknown\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo reports of PACS integration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFeasibility studies needed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCost-Effectiveness\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e? Unknown\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo economic analyses conducted\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHealth economics studies needed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTraining Requirements\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e? Limited\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTraining data sizes small; transfer learning potential\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFurther investigation needed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMaintenance \u0026amp; Updates\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e? Unknown\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo information on model drift/degradation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMonitoring protocols needed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eOverall Readiness Level: PROTOTYPE/RESEARCH\u003c/b\u003e (not ready for independent clinical use)\u003c/p\u003e \u003cp\u003e \u003cb\u003ePrerequisites for Clinical Integration\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eExternal validation on multi-institutional datasets\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eProspective clinical outcomes study\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eRegulatory approval (FDA/CE/national equivalent)\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eHealth economics evaluation\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eClinician training and credentialing protocols\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e"},{"header":"CONCLUSIONS","content":"\u003cp\u003eDeep learning-based artificial intelligence demonstrates considerable promise for objective assessment of mandibular third molar impaction, evaluation of anatomical relationships with the mandibular canal, and prediction of surgical complexity. Diagnostic accuracies reported across the literature (ranging from 78.91% to 99.0% for various applications) frequently exceed those of dental students and approach or match those of experienced oral surgeons.\u003c/p\u003e \u003cp\u003eHowever, substantial methodological limitations constrain the translational potential of current evidence. Single-institution dataset derivation, limited demographic diversity, reliance on cross-validation without independent external validation, and absence of prospective clinical outcomes data all limit the evidence strength and clinical applicability. These limitations are not unique to third molar impaction research but reflect broader challenges in artificial intelligence-assisted diagnostic imaging across medical specialties.\u003c/p\u003e \u003cp\u003eBefore integration into routine clinical practice, several critical gaps require attention: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) prospective, multi-institutional validation on diverse patient populations and imaging equipment configurations; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) prospective clinical outcomes studies examining the impact of artificial intelligence-assisted decision-making on surgical performance, complication rates, and patient satisfaction; (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) development and refinement of explainable artificial intelligence approaches to enhance clinician understanding and trust; (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) standardisation of outcome definitions and performance metrics to facilitate comparison across studies and enable meta-analysis; and (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) investigation of optimal implementation strategies and human-artificial intelligence interaction models.\u003c/p\u003e \u003cp\u003eCollaborative research involving oral radiologists, clinicians, computer scientists, and biostatisticians is essential to address these gaps and translate promising diagnostic accuracy into genuine clinical benefit. When these methodological and clinical validation steps are completed, artificial intelligence-assisted assessment of mandibular third molar impaction may substantially improve the standardisation, objectivity, and consistency of preoperative evaluation, enhance surgical planning and patient counselling, and reduce interobserver variability across the discipline.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAuthor Name\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eContribution\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDr. Muhammadh Munthasir\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eConception and design\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDr. Muhammadh Munthasir\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLiterature search and screening (abstract and full-text)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDr. Muhammadh Munthasir, Dr. Prathibha G, Dr. Parimala Sagar\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eData extraction\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDr. Muhammadh Munthasir\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRisk of bias assessment\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDr. Muhammadh Munthasir, Dr. Prathibha G, Dr. Parimala Sagar\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eManuscript drafting and critical revision\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDr. Muhammadh Munthasir, Dr. Prathibha G, Dr. Parimala Sagar, Dr. Kavitha Prasad\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFinal approval and accountability\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cstrong\u003eEthical Clearance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines and recommendations from the Joanna Briggs Institute for systematic reviews. Ethical clearance was not required as this review analysed published literature only and involved no human subjects or animal studies.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding and Support\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo financial support or funding was received for this systematic review from any source.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Sharing Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData extraction forms, search strategies, quality assessment results, and PRISMA 2020 checklist completion summary are available upon request from the corresponding author. This review includes no original patient data, as it constitutes a synthesis of published literature.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAI Use Declaration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis manuscript was prepared using artificial intelligence tools for literature search assistance and data organization. All literature interpretation, quality assessment, synthesis, and conclusions were conducted independently by the authors. The manuscript has been reviewed and verified for accuracy by all authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy Registration:\u003c/strong\u003e This systematic review was prospectively registered with PROSPERO bearing the ID: CRD420261279517 and the review was conducted and reported in accordance with standardised PRISMA 2020 methodology.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSearch Period:\u003c/strong\u003e Electronic database searches were conducted in August 2024 via PubMed (MEDLINE), Scopus, Web of Science, Google Scholar and IEEE Xplore. Studies published through January 2025 were included in the review.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNumber of Included Studies:\u003c/strong\u003e 15 peer-reviewed studies (published 2017-2024)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy Design:\u003c/strong\u003e Systematic review of diagnostic accuracy studies of artificial intelligence applications in third molar impaction assessment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eKey Findings Summary\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis systematic review synthesises evidence from 15 contemporary studies on the diagnostic performance of artificial intelligence for mandibular third molar impaction assessment. The findings demonstrate that deep learning-based models, particularly convolutional neural networks (CNNs), achieve impressive diagnostic accuracies:\u003c/p\u003e\n\u003cul class=\"decimal_type\"\u003e\n \u003cli\u003e\u003cstrong\u003eThird molar impaction classification:\u003c/strong\u003e 78.91% to 95.48% (median 89.2%)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eMandibular canal proximity assessment:\u003c/strong\u003e 72.32% to 99.0% (median 95.0%)\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eComparative performance:\u003c/strong\u003e AI models frequently exceed diagnostic accuracy of dental students and approach or match performance of experienced oral surgeons\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eDespite promising diagnostic accuracy, substantial methodological limitations constrain clinical applicability, including single-institution dataset derivation, limited demographic diversity, and absence of prospective clinical outcomes data. Future research should prioritise multi-institutional validation and prospective clinical outcomes studies before routine clinical implementation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eManuscript Highlights\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e✓ Comprehensive systematic review of 15 studies on artificial intelligence for third molar impaction assessment\u003cbr\u003e\u0026nbsp;✓ Detailed analysis of diagnostic performance across three application domains (impaction classification, canal proximity assessment, extraction difficulty prediction)\u003cbr\u003e\u0026nbsp;✓ Methodological quality assessment using adapted Joanna Briggs Institute tools\u003cbr\u003e\u0026nbsp;✓ Identification of critical evidence gaps and recommendations for future research\u003cbr\u003e\u0026nbsp;✓ Clear clinical guidance on current appropriate use of AI as adjunctive decision-support tool\u003cbr\u003e\u0026nbsp;✓ Discussion of implementation strategies and validation requirements for routine clinical practice\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eManuscript Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eI/We certify that this work is original and has not been published or submitted for publication elsewhere in any form. All authors have read and approved the final manuscript and accept responsibility for its contents. The work conforms to the ethical guidelines and policies of the Indian Journal of Medical Research.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDate:\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;16/01/2026\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cimg width=\"369\" height=\"169\" src=\"data:image/jpeg;base64,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\" alt=\"IMG-20260121-WA0013\"\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCorresponding Author Contact:\u003c/strong\u003e\u003cbr\u003e\u0026nbsp;Email: [email protected]\u003cbr\u003e\u0026nbsp;Phone: +91 9677100201\u003c/p\u003e\u003ch2\u003eACKNOWLEDGEMENTS\u003c/h2\u003e\n\u003cp\u003eThe authors acknowledge the contributions of librarians in formulating the search strategy and the authors of the included studies who provided additional methodological details when requested. No financial support or funding was received for this systematic review.\u003c/p\u003e\n\u003ch2\u003eCONFLICT OF INTERESTS\u003c/h2\u003e\n\u003cp\u003eThe authors declare no conflict of interests.\u003c/p\u003e\n\u003ch2\u003eDATA SHARING STATEMENT\u003c/h2\u003e\n\u003cp\u003eData extraction forms, search strategies, and quality assessment results are available upon request from the corresponding author. No original patient data are included in this review.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBorle RM. Textbook of oral and maxillofacial surgery. 3rd ed. Jaypee Brothers Medical; 2014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSujon MK et al. (2022). Global prevalence of impacted third molars: A systematic review and meta-analysis. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 134(5), 512\u0026ndash;525.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlfadil F, Almajed A. Classification systems for impacted mandibular third molars: Clinical application and complications. J Oral Maxillofac Surg. 2020;78(12):2087\u0026ndash;98.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSwift JQ, Nelson RE. Wisdom teeth and their impact on orthodontic treatment. Dental Clin N Am. 2012;56(2):307\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHung K, Montalvao C, Tanaka R, Kawai T, Bornstein MM. The use and performance of artificial intelligence applications in dental and maxillofacial radiology: A systematic review. Dentomaxillofacial Radiol. 2019;48(7):20190107.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePage MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ. 2021;372:n71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFaadiya AN, Widyaningrum R, Arindra PK, Diba SF. The diagnostic performance of impacted third molars in the mandible: A review of deep learning on panoramic radiographs. Saudi Dent J. 2024;36:404\u0026ndash;12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCelik C. Deep learning-based classification of mandibular third molar impaction severity. Oral Surgery, Oral Medicine. Oral Pathol Oral Radiol. 2022;134(3):312\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChoi SY, Park JW, Kim SJ, et al. Diagnostic performance of convolutional neural networks for mandibular canal proximity assessment. J Oral Maxillofac Surg. 2022;80(11):1847\u0026ndash;56.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFukuda M, Inamoto K, Shibata N, et al. Artificial intelligence in oral and maxillofacial imaging: Current applications and limitations. Oral Surgery, Oral Medicine. Oral Pathol Oral Radiol. 2020;130(1):72\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKwon SJ, et al. Prediction of extraction time for impacted mandibular third molars using hybrid deep learning models. Br J Oral Maxillofac Surg. 2022;60(4):412\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSukegawa S, Yoshii K, Hattori T, et al. Deep learning-based image analysis for automated classification of mandibular third molar impaction. Imaging Sci Dentistry. 2022;52(2):189\u0026ndash;98.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoo HG, Park JS, Kim YH. (2021). Convolutional neural network application for assessing surgical difficulty of impacted third molars. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 131(4), 412\u0026ndash;422.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu X, Chen L, Wang Y. Automated segmentation and classification of mandibular third molars using deep convolutional neural networks. J Dent Res. 2021;100(5):525\u0026ndash;33.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEkert T, Krois J, Meinhold L, et al. Deep learning for the radiographic diagnosis of periodontal disease. J Periodontol. 2019;90(12):1459\u0026ndash;72.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVinayahalingam S, et al. Classification of permanent teeth in panoramic radiographs using convolutional neural networks. Comput Biol Med. 2021a;135:104551.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVinayahalingam S, et al. Automated detection of caries using deep learning and its impact on clinical workflow. Oral Surgery, Oral Medicine. Oral Pathol Oral Radiol. 2021b;131(4):445\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEstai M, et al. Deep learning approaches for automated tooth detection in digital dental images. J Dent Res. 2022;101(5):578\u0026ndash;87.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVranckx M, Marechal C, Lacko M, et al. Machine learning prediction of third molar eruption status based on panoramic radiographs. Clin Oral Invest. 2020;24(9):3047\u0026ndash;58.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMandeel A, Sharif S, Ahmed A. (2022). Deep learning for pneumonia detection on chest radiographs: A systematic review. Radiology: Artif Intell, 4(1), e210118.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePratiwi B, Setiawan E, Harefa R, Safitri V. Deep neural networks for skin lesion classification: A systematic review. J Dermatological Treat. 2021;32(2):113\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMurata M, Nishii Y, Nakamori K et al. (2019). Maxillary sinus classification and volumetric analysis using artificial intelligence. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 128(1), 45\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee JH, Kim DH, Jeong SN, Choi SH. Diagnosis and prediction of periodontally compromised teeth using a deep learning-based convolutional neural network algorithm. J Periodontal Implant Sci. 2018;48(2):114\u0026ndash;23.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWayland HE. Surgical management of impacted teeth. In: Kamadjaja DB, Morkel RE, editors. Oral and maxillofacial surgery for the dental student. Springer; 2018. pp. 287\u0026ndash;312.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNagaraj K, Sharma Y, Das S. Role of cone-beam computed tomography in assessment of mandibular third molar impaction. J Oral Maxillofac Surg. 2016;74(8):1562\u0026ndash;70.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmisha A, Malik P, Pathania M, Rathaur VK. Overview of artificial intelligence in medicine. J Family Med Prim Care. 2019;8(7):2328\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYasa M, Cosar M, Karakol B. Computer-aided diagnosis systems in dentistry: Current status and future perspectives. Int J Oral Sci. 2021;13:18.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8705343/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8705343/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eContext: Mandibular third molar impaction presents significant diagnostic and surgical challenges. Artificial intelligence, particularly deep learning approaches, has emerged as a promising tool for classification, risk assessment, and surgical planning in oral and maxillofacial surgery.\u003c/p\u003e \u003cp\u003eObjective: To systematically evaluate the diagnostic performance and clinical applicability of cognitive computing techniques for assessing mandibular third molar impaction and associated surgical complexities.\u003c/p\u003e \u003cp\u003eEvidence Acquisition: A systematic literature search was conducted in accordance with PRISMA 2020 guidelines using PubMed, Scopus, Google Scholar Web of Science, and IEEE Xplore databases for studies published up to January 2025. Search terms included combinations of \"third molar,\" \"wisdom tooth,\" \"impaction,\" \"artificial intelligence,\" \"deep learning,\" \"radiograph,\" and \"convolutional neural networks.\" Two reviewers independently screened and extracted data. Risk of bias was assessed using adapted Joanna Briggs Institute (JBI) quality assessment tools.\u003c/p\u003e \u003cp\u003eResults: Fifteen studies met inclusion criteria, involving primarily convolutional neural networks (CNNs), support vector machines (SVMs), and ensemble classifiers. Most studies utilised panoramic radiographs, whilst some incorporated cone-beam computed tomography (CBCT) or clinical metadata. AI systems demonstrated diagnostic accuracies ranging from 78.91% to 99.0% for impaction classification, and 72.32% to 99.0% for mandibular canal proximity assessment. Multiple studies showed that AI models outperformed dental students and general practitioners, though performance varied across different architectural approaches and training datasets. Several models demonstrated potential for predicting extraction time (mean absolute error\u0026thinsp;\u0026lt;\u0026thinsp;3 minutes) and postoperative complications (accuracy up to 98%).\u003c/p\u003e \u003cp\u003eLimitations: Limited generalisability due to single-institution datasets, variable dataset sizes, and heterogeneous methodological approaches. Publication bias and selective outcome reporting were not comprehensively assessed.\u003c/p\u003e \u003cp\u003eConclusions: Deep learning-based artificial intelligence demonstrates considerable promise for objective classification of mandibular third molar impaction, proximity assessment with the mandibular canal, and prediction of surgical complexity. However, further development using large multi-institutional datasets, adherence to standardised diagnostic test protocols, and prospective clinical validation are essential before integration into routine clinical practice. Collaborative research involving oral radiologists, clinicians, and computer scientists is required to optimise AI model development and ensure clinical applicability.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e","manuscriptTitle":"Cognitive Computing Approaches for Assessing Mandibular Third Molar Impaction: A Systematic Review","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-29 10:40:54","doi":"10.21203/rs.3.rs-8705343/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f85f5421-c7a9-4de2-bdff-1f640b12efbc","owner":[],"postedDate":"January 29th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-07T06:09:32+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-29 10:40:54","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8705343","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8705343","identity":"rs-8705343","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}