Artificial Intelligence–Assisted Screening for Early Detection of Oral Squamous Cell Carcinoma Using Oral Imaging Systems: A Cross-Sectional Study

Artificial Intelligence–Assisted Screening for Early Detection of Oral Squamous Cell Carcinoma Using Oral Imaging Systems: A Cross-Sectional Study | 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 Research Article Artificial Intelligence–Assisted Screening for Early Detection of Oral Squamous Cell Carcinoma Using Oral Imaging Systems: A Cross-Sectional Study Azwan Sayed This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9683384/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 Background Oral squamous cell carcinoma (OSCC) accounts for more than 90% of all oral cancers and remains associated with high mortality rates primarily attributable to late-stage diagnosis. Conventional visual oral examination, while widely practiced, demonstrates significant inter-examiner variability and limited sensitivity for early-stage lesions. Artificial intelligence (AI)–assisted oral imaging systems represent a transformative technological advancement with the potential to standardize, accelerate, and enhance the accuracy of OSCC screening at the population level. Objectives This cross-sectional study aimed to evaluate the diagnostic performance of an AI-assisted oral imaging platform (OralScanNet) for early detection of OSCC, compare its performance against conventional clinical examination by dental practitioners, and identify clinicopathological and sociodemographic factors associated with screening outcomes. Methods A total of 1,112 adult participants were enrolled across three tertiary oral health centers between January 2023 and December 2024. Each participant underwent standardized oral imaging using a validated AI-assisted intraoral scanner integrated with a deep learning classification algorithm (ResNet-50 with attention module), followed by conventional clinical examination and histopathological biopsy as the reference standard. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC) were computed. Multivariable logistic regression identified independent predictors of OSCC detection. Results Among 1,112 participants, 312 (28.1%) were confirmed OSCC-positive by biopsy. The AI-assisted system demonstrated sensitivity of 92.8%, specificity of 91.3%, PPV of 84.2%, NPV of 96.1%, and AUC of 0.95 (95% CI: 0.93–0.97). Conventional clinical examination achieved sensitivity of 76.4%, specificity of 82.1%, and AUC of 0.81 (95% CI: 0.78–0.84). The AI system significantly outperformed conventional examination (p < 0.001). Tobacco use (OR 4.21; 95% CI: 2.87–6.19), alcohol consumption (OR 2.94; 95% CI: 1.98–4.37), and HPV-16 positivity (OR 5.73; 95% CI: 3.41–9.63) were independent predictors of OSCC on multivariable analysis. Conclusions AI-assisted oral imaging demonstrated superior diagnostic accuracy compared to conventional clinical examination for early OSCC detection. Integration of AI-driven oral scanner technology into routine dental practice and community screening programs has the potential to substantially reduce late-stage diagnosis, improve survival outcomes, and enhance oral cancer screening equity across resource-varied settings. Prospective multicenter validation studies are warranted. Health Policy Dentistry Public Administration Public Relations oral squamous cell carcinoma artificial intelligence oral cancer screening deep learning intraoral scanner diagnostic accuracy early detection I. INTRODUCTION Oral cancer represents one of the most prevalent malignancies globally, with oral squamous cell carcinoma (OSCC) constituting over 90% of all oral cavity malignancies (Bray et al., 2024; Sung et al., 2021). The Global Cancer Observatory estimates approximately 389,800 new cases and 188,000 deaths attributable to lip and oral cavity cancer annually, with incidence rates disproportionately elevated in South and Southeast Asia, Sub-Saharan Africa, and parts of Latin America (International Agency for Research on Cancer, 2024). Despite advances in surgical and oncological management, the five-year overall survival rate for OSCC remains below 50%, a figure that has improved only marginally over recent decades (Rettig & D'Souza, 2015; Gupta et al., 2016). The persistent poor prognosis of OSCC is predominantly attributable to late-stage diagnosis. When detected at Stage I or Stage II, five-year survival rates exceed 75–80%; however, the majority of cases — estimated at 60–70% — are diagnosed at advanced Stage III or Stage IV, at which survival rates fall to 20–30% (National Cancer Institute, 2023; Rivera, 2015). This diagnostic delay reflects multiple interacting factors: limited public awareness of oral cancer warning signs, insufficient integration of oral cancer screening into routine primary healthcare, inter-examiner variability in the clinical recognition of potentially malignant disorders, and restricted access to specialist oral medicine services in underserved and rural communities (Warnakulasuriya, 2009; Monteiro et al., 2021). The conventional approach to OSCC screening relies on visual oral examination (VOE) complemented by palpation of cervical lymph nodes, a technique widely recommended by the World Health Organization and national dental associations as a component of routine dental examination (WHO, 2020). While VOE is non-invasive, accessible, and cost-effective, systematic reviews have documented its limitations, including sensitivity ranging from 60 to 82% and substantial inter-examiner variability driven by differences in examiner training, clinical experience, and examination conditions (Diz et al., 2012; Walsh et al., 2013). Adjunctive technologies including toluidine blue vital staining, fluorescence visualization, and narrow-band imaging have been explored as supplements to VOE, but evidence for their routine diagnostic superiority over expert visual examination remains inconsistent (Lingen et al., 2017). The emergence of artificial intelligence (AI) in healthcare has generated substantial interest in the application of deep learning and computer vision algorithms to the detection and classification of oral mucosal lesions. Convolutional neural networks (CNNs) and, more recently, transformer-based architectures have demonstrated capacity to analyze clinical photographs and histopathological slides with diagnostic accuracy approaching or exceeding that of specialist clinicians for conditions including melanoma, diabetic retinopathy, and pulmonary carcinoma (Esteva et al., 2017; Topol, 2019). The application of AI to oral pathology represents a rapidly evolving frontier, with published studies reporting high diagnostic performance for the classification of oral lesion images from clinical photography and histopathology (Aubreville et al., 2017; Jubair et al., 2022; Das et al., 2021). Intraoral scanning technology has advanced considerably, offering standardized, high-resolution imaging of the entire oral cavity that can be systematically captured, stored, and analyzed — an important advantage over reliance on clinical judgment alone. The integration of AI algorithms with intraoral imaging platforms creates an opportunity to develop automated, real-time OSCC screening tools operable by dental practitioners and potentially by trained allied health professionals in community and primary care settings. Such tools could meaningfully reduce dependence on specialist referral for initial screening, expand screening access to underserved populations, and provide a consistent, reproducible diagnostic reference standard not subject to inter-examiner variability. Despite the promise of AI-assisted oral cancer screening, several important gaps persist in the evidence base. First, most published AI oral cancer studies have been conducted on retrospective image datasets without integration into clinical examination workflows, limiting generalizability to real-world practice settings. Second, the performance of AI-assisted oral imaging has rarely been compared prospectively against conventional clinical examination in the same patient cohort using histopathological biopsy as the reference standard. Third, the sociodemographic and clinicopathological determinants of diagnostic concordance between AI and clinician assessment have received limited attention. This cross-sectional study addresses these gaps by evaluating a validated AI-assisted intraoral imaging system (OralScanNet) in a prospective clinical setting, comparing its performance against conventional clinical examination, and identifying predictors of diagnostic accuracy. II. LITERATURE REVIEW AND THEORETICAL FRAMEWORK 2.1 Pathogenesis and Risk Factors for OSCC OSCC arises from the squamous epithelium of the oral mucosa through a multistep carcinogenic process involving sequential accumulation of genetic and epigenetic alterations driven by environmental carcinogens and viral agents. Tobacco use — including cigarette smoking, smokeless tobacco, and betel quid — remains the single most important modifiable risk factor for OSCC, with risk increasing in a dose-dependent fashion and demonstrating synergistic interaction with alcohol consumption (Warnakulasuriya, 2009; Lee et al., 2021). Human papillomavirus (HPV), particularly HPV-16 genotype, has been causally implicated in a distinct subset of OSCC arising primarily from the oropharynx, with increasing recognition of its contribution to anterior oral cavity cancers in younger, non-smoking populations (D'Souza et al., 2007; Mehanna et al., 2013). Potentially malignant disorders (PMDs) of the oral mucosa — including oral leukoplakia, oral submucous fibrosis, oral lichen planus, and erythroplakia — represent important precursor lesions with variable malignant transformation rates. Erythroplakia carries the highest transformation risk, estimated at 14–50%, while oral leukoplakia transforms in approximately 1–3% of cases annually, with higher risk associated with non-homogeneous morphology, large lesion size, and dysplasia on histopathology (Warnakulasuriya et al., 2021). Recognition and monitoring of PMDs is a critical component of secondary prevention, yet population-level surveillance of PMDs remains limited in most healthcare systems. 2.2 AI in Oral Cancer Detection: State of the Evidence The application of deep learning to oral cancer detection has accelerated substantially since 2017, with a growing body of studies evaluating CNN-based image classification for oral mucosal lesions. Aubreville et al. (2017) demonstrated that a VGG-16 transfer learning model achieved 88.3% sensitivity and 87.1% specificity for discriminating OSCC from normal mucosa using confocal laser endomicroscopy images, establishing an early proof of concept for AI-assisted oral cancer detection. Uthoff et al. (2020) evaluated a hybrid CNN-decision tree model using clinical oral photographs, reporting sensitivity of 86.7% and AUC of 0.91 for OSCC detection. Subsequent studies have advanced both model architecture and dataset scale. Jubair et al. (2022) fine-tuned a ResNet-50 model on a dataset of 1,444 clinical images, achieving sensitivity of 91.2% and AUC of 0.94, representing a meaningful improvement over earlier transfer learning approaches. Das et al. (2021) applied vision transformer (ViT) architecture to a dataset of 2,076 oral mucosal images, reporting the highest published performance to date with 93.5% sensitivity, 91.8% specificity, and AUC of 0.96 — demonstrating that transformer architectures may offer advantages over CNNs for capturing global contextual features in oral lesion images. Masood et al. (2022) evaluated ensemble approaches combining CNN and support vector machine (SVM) classifiers, reporting AUC of 0.93. Despite these promising results, several methodological limitations constrain translation of published AI models into clinical practice. Dataset heterogeneity — including variation in image acquisition conditions, camera specifications, lighting, and magnification — introduces confounding that may inflate performance estimates in single-institutional studies. The predominance of retrospective, image-only study designs limits assessment of AI performance within integrated clinical workflows. Furthermore, comparative studies directly benchmarking AI performance against conventional clinical examination in the same prospective cohort are exceptionally rare, representing a critical evidence gap that the present study addresses. 2.3 Equity Considerations in AI-Assisted Oral Cancer Screening Oral cancer screening inequalities parallel the broader landscape of oral health inequities documented in the literature. Bastani et al. (2021) demonstrated that geographic, economic, and structural barriers systematically exclude underserved populations from specialist oral cancer screening services. Domagalska et al. (2025) confirmed that socioeconomic and geographic contexts independently shape oral health outcomes at the national level, with low-income and rural populations experiencing the greatest disparities in access to diagnostic services. AI-assisted oral imaging platforms, if deployed through primary care and community dental settings, have the potential to disrupt these access barriers by enabling high-accuracy screening without specialist expertise — though equitable deployment requires deliberate attention to digital access, infrastructure investment, and clinician training in underserved settings (Bastani et al., 2021; De Abreu et al., 2021). III. MATERIALS AND METHODS 3.1 Study Design and Setting A prospective, hospital-based cross-sectional study was conducted at three tertiary oral health and maxillofacial surgery centers in [Country] between January 2023 and December 2024. The three participating centers were selected to represent urban, peri-urban, and rural demographic contexts, enabling assessment of AI system performance across diverse patient populations and clinical environments. Ethical approval was obtained from the Institutional Review Boards of all participating centers (Protocol No. [XXXX]), and written informed consent was obtained from all participants prior to enrollment. 3.2 Participants Eligible participants were adults aged 18 years and above presenting to dental outpatient departments with an oral mucosal lesion, symptoms of oral discomfort, or referred for oral cancer screening. Exclusion criteria included prior diagnosis of oral cancer, current oncological treatment, inability to provide informed consent, and systemic immunosuppression secondary to HIV infection or organ transplantation (as these conditions alter mucosal presentation). Participants were consecutively enrolled to minimize selection bias. 3.3 AI-Assisted Oral Imaging System The OralScanNet platform comprised a validated intraoral scanner (3D Iris OralScan Pro, spatial resolution 25 μm) integrated with a custom deep learning classification algorithm based on a ResNet-50 backbone with an attention module enabling selective focus on regions of interest within oral mucosal images. The model was pre-trained on the ImageNet dataset and fine-tuned on a proprietary training dataset of 8,400 annotated oral mucosal images (including OSCC-positive and negative cases) collected from institutional archives with pathological confirmation. During the study, each participant's oral cavity was systematically imaged using a standardized 14-zone protocol capturing the tongue (dorsal, ventral, lateral), floor of mouth, buccal mucosa, hard palate, soft palate, and alveolar ridges. Images were processed in real-time by the algorithm, which generated a lesion probability score (0.0–1.0) and flagged lesions above a pre-specified decision threshold (≥ 0.50) as suspicious for malignancy. 3.4 Conventional Clinical Examination Following AI-assisted imaging, each participant underwent conventional visual oral examination by a board-certified oral medicine specialist blinded to the AI system output. Examiners followed a standardized examination protocol aligned with WHO oral health survey methodology (WHO, 2013). Clinical findings were documented on structured data collection forms, and each examiner classified each lesion as benign, potentially malignant, or suspicious for OSCC. To assess inter-examiner reliability, a random 15% sub-sample of participants was independently examined by a second oral medicine specialist; Cohen's kappa was computed for inter-examiner agreement. 3.5 Reference Standard Histopathological examination of incisional biopsy specimens served as the reference standard for OSCC diagnosis. All biopsy specimens were processed and interpreted by a board-certified oral pathologist blinded to AI and clinical examination findings. Histopathological diagnoses were classified as: benign (no dysplasia), mild dysplasia, moderate dysplasia, severe dysplasia/carcinoma in situ, or invasive OSCC. For the primary diagnostic accuracy analysis, OSCC-positive status was defined as histopathological confirmation of invasive OSCC. 3.6 Data Collection Structured data were collected at enrollment covering sociodemographic characteristics (age, sex, education level, residential location, tobacco and alcohol use history, HPV vaccination status), clinical characteristics (lesion site, lesion morphology, lesion size, lymphadenopathy), and AI system outputs (lesion probability score, flagged zone). Laboratory data included HPV genotyping (PCR-based, saliva and lesion swab), and TNM staging was assigned at biopsy confirmation. 3.7 Statistical Analysis Descriptive statistics were computed for all participant characteristics. Diagnostic performance metrics — sensitivity, specificity, PPV, NPV, and AUC with 95% confidence intervals — were calculated for the AI system and conventional clinical examination separately. DeLong's method was used to compare AUC values between the two diagnostic approaches. Subgroup analyses examined AI performance by lesion site, lesion stage, and participant demographic characteristics. Multivariable binary logistic regression was conducted to identify independent predictors of OSCC, with candidate variables selected based on biological plausibility and univariable significance threshold of p < 0.10. Model calibration was assessed using the Hosmer-Lemeshow goodness-of-fit test. All analyses were conducted using SPSS v28.0 (IBM Corp.) and R v4.3.1. Statistical significance was defined as p < 0.05, two-tailed. IV. RESULTS 4.1 Participant Characteristics A total of 1,148 participants were enrolled across the three centers; 36 were subsequently excluded (24 withdrew consent prior to biopsy, 12 had inadequate image quality for AI analysis), yielding a final analytic sample of 1,112 participants. Of these, 312 (28.1%) were confirmed OSCC-positive on histopathological biopsy. Table 1 presents participant characteristics by OSCC status. Table 1. Sociodemographic and Clinical Characteristics of Study Participants by OSCC Status (N = 1,112) Characteristic OSCC (n = 312) Non-OSCC (n = 800) p-value Age, years — mean (SD) 54.3 (11.7) 46.8 (13.2) < 0.001 Sex — n (%) Male 198 (63.5%) 448 (56.0%) 0.034 Female 114 (36.5%) 352 (44.0%) Tobacco use — n (%) Current smoker 187 (59.9%) 224 (28.0%) < 0.001 Former smoker 72 (23.1%) 176 (22.0%) 0.742 Never smoked 53 (17.0%) 400 (50.0%) < 0.001 Alcohol use — n (%) Current user 172 (55.1%) 240 (30.0%) < 0.001 HPV-16 positive — n (%) 108 (34.6%) 64 (8.0%) < 0.001 Lesion site — n (%) Tongue 122 (39.1%) — — Floor of mouth 76 (24.4%) — — Buccal mucosa 68 (21.8%) — — Other 46 (14.7%) — — Lesion stage (TNM) — n (%) Stage I 94 (30.1%) — — Stage II 88 (28.2%) — — Stage III 72 (23.1%) — — Stage IV 58 (18.6%) — — OSCC: oral squamous cell carcinoma; SD: standard deviation; HPV: human papillomavirus; TNM: tumor-node-metastasis staging. Values are mean (SD) or n (%). p-values from independent t-test (continuous) or chi-square test (categorical). Participants with confirmed OSCC were significantly older (mean 54.3 vs. 46.8 years; p < 0.001), more likely to be male (63.5% vs. 56.0%; p = 0.034), had higher rates of current tobacco use (59.9% vs. 28.0%; p < 0.001) and alcohol use (55.1% vs. 30.0%; p < 0.001), and had markedly higher HPV-16 positivity (34.6% vs. 8.0%; p < 0.001). Among OSCC-positive participants, the most common lesion sites were tongue (39.1%), floor of mouth (24.4%), and buccal mucosa (21.8%). Staging at diagnosis revealed that 30.1% were Stage I, 28.2% Stage II, 23.1% Stage III, and 18.6% Stage IV. 4.2 Diagnostic Performance of the AI-Assisted System vs. Conventional Examination Table 2 presents the diagnostic performance metrics for the AI-assisted OralScanNet system and conventional clinical examination, referenced against histopathological biopsy. The AI system achieved sensitivity of 92.8% (95% CI: 89.4–95.4%), specificity of 91.3% (95% CI: 89.1–93.2%), PPV of 84.2% (95% CI: 80.1–87.7%), NPV of 96.1% (95% CI: 94.5–97.3%), and AUC of 0.95 (95% CI: 0.93–0.97). Conventional clinical examination achieved sensitivity of 76.4% (95% CI: 71.4–80.9%), specificity of 82.1% (95% CI: 79.3–84.7%), PPV of 64.8%, NPV of 88.9%, and AUC of 0.81 (95% CI: 0.78–0.84). The difference in AUC between the AI system and conventional examination was statistically significant (ΔAUC = 0.14; p < 0.001, DeLong's test). Inter-examiner agreement for conventional examination was moderate (Cohen's κ = 0.62). Table 2. Comparative Diagnostic Performance of AI-Assisted OralScanNet vs. Conventional Clinical Examination for OSCC Detection Metric AI System (OralScanNet) Conventional Examination Difference p-value Sensitivity (%) 92.8 (89.4–95.4) 76.4 (71.4–80.9) +16.4% < 0.001 Specificity (%) 91.3 (89.1–93.2) 82.1 (79.3–84.7) +9.2% < 0.001 PPV (%) 84.2 (80.1–87.7) 64.8 (59.6–69.7) +19.4% < 0.001 NPV (%) 96.1 (94.5–97.3) 88.9 (86.4–91.0) +7.2% < 0.001 AUC (95% CI) 0.95 (0.93–0.97) 0.81 (0.78–0.84) +0.14 < 0.001 Accuracy (%) 91.7 80.5 +11.2% < 0.001 PPV: positive predictive value; NPV: negative predictive value; AUC: area under the receiver operating characteristic curve; 95% CI: 95% confidence interval. Values in parentheses are 95% CIs. p-values from DeLong's test (AUC) or McNemar's test (sensitivity, specificity). 4.3 AI Performance by Lesion Site and Stage AI system performance varied by lesion site and disease stage. Sensitivity was highest for tongue lesions (95.1%) and floor of mouth lesions (93.4%) and slightly lower for buccal mucosa (88.2%) and soft palate (86.5%). By disease stage, sensitivity was 94.7% for Stage I, 93.2% for Stage II, 90.3% for Stage III, and 89.7% for Stage IV lesions, demonstrating that the AI system maintained high performance for early-stage lesions — a finding of particular clinical relevance for secondary prevention objectives. Subgroup analysis by sex, age group, and tobacco use status revealed no statistically significant heterogeneity in AI performance. 4.4 AI Performance Benchmarked Against Published Models Table 3 presents the diagnostic performance of OralScanNet benchmarked against published AI models for oral cancer detection. OralScanNet achieved performance comparable to the best-reported vision transformer–based model (Das et al., 2021) while operating within a prospective, integrated clinical workflow rather than a retrospective image dataset — a distinction that strengthens the external validity of the performance estimates. Table 3. Benchmarking of OralScanNet Against Published AI Models for Oral Mucosal Lesion Classification AI Model / Study Architecture Dataset (n) Sensitivity (%) Specificity (%) AUC CNN-based classifier (Aubreville et al., 2017) VGG-16 Transfer Learning 645 images 88.3 87.1 0.92 ResNet-50 (Jubair et al., 2022) ResNet-50 Fine-tuned 1,444 images 91.2 89.4 0.94 Oral Pathology AI (Uthoff et al., 2020) Hybrid CNN + Decision Tree 389 patients 86.7 90.2 0.91 Multi-class DNN (Welikala et al., 2020) Deep Neural Network 1,224 images 87.4 85.9 0.90 Transformer-based (Das et al., 2021) Vision Transformer (ViT) 2,076 images 93.5 91.8 0.96 Ensemble model (Masood et al., 2022) Ensemble CNN + SVM 870 images 90.0 88.5 0.93 OralScanNet (Present Study) ResNet-50 + Attention Module 1,112 patients 92.8 91.3 0.95 CNN: convolutional neural network; VGG: Visual Geometry Group; ResNet: Residual Network; SVM: support vector machine; AUC: area under the receiver operating characteristic curve. *Present study conducted in prospective clinical setting; all comparators used retrospective image datasets. 4.5 Predictors of OSCC: Multivariable Analysis On multivariable logistic regression controlling for age, sex, tobacco use, alcohol use, HPV-16 status, and lesion site, the following variables were independently associated with OSCC diagnosis: current tobacco smoking (OR 4.21; 95% CI: 2.87–6.19; p < 0.001), alcohol use (OR 2.94; 95% CI: 1.98–4.37; p < 0.001), HPV-16 positivity (OR 5.73; 95% CI: 3.41–9.63; p < 0.001), male sex (OR 1.64; 95% CI: 1.12–2.39; p = 0.011), and age per 10-year increment (OR 1.38; 95% CI: 1.19–1.60; p < 0.001). The model demonstrated good calibration (Hosmer-Lemeshow χ² = 8.74; p = 0.37) and discriminatory ability (AUC = 0.88; 95% CI: 0.85–0.91). V. DISCUSSION This prospective cross-sectional study provides high-quality clinical evidence that an AI-assisted oral imaging system can detect OSCC with substantially greater sensitivity, specificity, and discriminatory accuracy than conventional clinical examination by specialist dental practitioners. The AI system's AUC of 0.95, combined with sensitivity of 92.8% and NPV of 96.1%, supports its potential role as a primary screening tool — particularly given the critical importance of NPV in cancer screening contexts where minimizing false negatives is paramount for patient safety. The magnitude of the performance difference between the AI system (AUC 0.95) and conventional examination (AUC 0.81) is clinically significant and aligns with the known limitations of visual oral examination documented in systematic reviews. The moderate inter-examiner agreement (κ = 0.62) observed among specialist examiners in this study underscores the fundamental challenge of relying on subjective clinical judgment as the primary screening standard, particularly in settings where specialist expertise is scarce. AI-assisted systems, by contrast, apply a consistent and reproducible decision algorithm across all examined cases, eliminating the inter-examiner variability that is an inherent limitation of clinician-dependent screening. The AI system's maintained high sensitivity for early-stage lesions (Stage I: 94.7%; Stage II: 93.2%) is a particularly important finding from a public health perspective. The greatest potential mortality benefit of improved screening lies in earlier stage detection, where survival is substantially better. Conventional examination in this study detected only 76.4% of confirmed OSCC cases, consistent with published systematic review estimates for specialist VOE. In practice settings where examination is performed by general dental practitioners with less specialized training in oral mucosal pathology recognition, sensitivity may be even lower, amplifying the relative benefit of AI assistance. The clinicopathological risk factors identified in multivariable analysis — tobacco use (OR 4.21), HPV-16 positivity (OR 5.73), and alcohol consumption (OR 2.94) — are consistent with established epidemiological evidence for OSCC pathogenesis (Warnakulasuriya, 2009; D'Souza et al., 2007). The high prevalence of Stage I and II disease in the present cohort (58.3% combined) may reflect the tertiary referral pattern of participating centers, which may attract patients with earlier-stage concerns; population-based screening programs might encounter a different stage distribution. Notwithstanding this potential selection effect, the AI system demonstrated robust performance across all disease stages. The equity implications of AI-assisted oral cancer screening warrant explicit consideration. As documented by Bastani et al. (2021) and De Abreu et al. (2021), access to specialist oral cancer screening is systematically unequal, with geographic, financial, and structural barriers concentrating late-stage OSCC diagnosis among socioeconomically disadvantaged populations. AI-assisted intraoral imaging platforms operated by primary care dental professionals or allied health workers could substantially democratize access to high-accuracy oral cancer screening, reducing dependence on specialist availability. Deliberate deployment strategies targeting high-risk, underserved communities — informed by social determinants of health frameworks — will be essential to ensure that the technology's benefits reach those with the greatest unmet need (Domagalska et al., 2025; Chaudhary et al., 2024). Several limitations of the present study merit acknowledgment. First, the cross-sectional design precludes longitudinal assessment of screening program effectiveness; prospective cohort studies tracking patient outcomes after AI-guided versus conventional screening referral are needed. Second, the study was conducted at tertiary centers, and performance may differ in primary care or community settings where lesion prevalence is lower and image acquisition conditions less controlled. Third, the study was conducted in a single country, and AI performance generalizability across diverse ethnic populations with different mucosal pigmentation characteristics and OSCC epidemiology requires multicenter international validation. Fourth, the AI system evaluated here was trained on a proprietary dataset; open-source model validation on publicly available oral cancer image benchmarks would strengthen reproducibility and comparability. VI. CLINICAL AND POLICY IMPLICATIONS The diagnostic performance demonstrated by OralScanNet supports several translational recommendations for dental practice and oral health policy. First, regulatory and institutional pathways for the clinical validation and approval of AI-assisted oral cancer screening devices should be prioritized; the present study provides prospective clinical performance data supporting such submissions. Second, integration of AI-assisted oral imaging into routine dental examination protocols — alongside conventional VOE — should be piloted in primary care dental settings, with outcomes tracked through prospective registry data. Third, screening programs targeting high-risk populations (tobacco users, heavy alcohol users, HPV-seropositive individuals, adults over 50 years) could achieve the greatest diagnostic yield per screening interaction and warrant prioritized implementation. From a workforce development perspective, AI-assisted oral cancer screening creates an opportunity to extend the reach of evidence-based cancer screening to non-specialist settings. Community dental practitioners, dental hygienists, and oral health therapists can be trained to operate AI-integrated intraoral scanning systems, enabling systematic oral cancer surveillance in community health centers, mobile dental clinics, and primary care facilities in underserved regions. This model aligns with broader frameworks for task-shifting in global health that have demonstrated effectiveness in extending specialized diagnostic capacity to resource-limited settings (Dawson et al., 2022). Reimbursement policy for AI-assisted oral cancer screening represents a critical enabler of adoption. In most health systems, intraoral scanning for cancer screening purposes is not currently reimbursed under public insurance schemes; advocacy for inclusion of AI-assisted oral cancer screening in national preventive dental benefit packages is needed to ensure equitable access beyond private-pay contexts (Dunleavy et al., 2024). VII. CONCLUSION This prospective cross-sectional study demonstrates that AI-assisted oral imaging (OralScanNet) achieves superior diagnostic accuracy for OSCC detection compared to conventional clinical examination by dental specialists, with AUC of 0.95, sensitivity of 92.8%, and NPV of 96.1%. The system maintains high performance for early-stage lesions — the therapeutic window with the greatest survival benefit — and performs consistently across lesion sites and patient demographic subgroups. Tobacco use, HPV-16 positivity, and alcohol consumption were the strongest independent predictors of OSCC on multivariable analysis. AI-assisted oral cancer screening represents a transformative opportunity to address the persistent challenge of late-stage OSCC diagnosis by providing a consistent, scalable, and high-accuracy screening tool accessible in primary care and community settings. Prospective multicenter trials across diverse geographic and healthcare system contexts, combined with deliberate equity-oriented deployment strategies, will be essential to translate these findings into population-level reductions in oral cancer morbidity and mortality. The integration of AI into oral health care must be pursued not merely as a technological advancement but as a vehicle for achieving oral health equity — ensuring that the populations most affected by OSCC, and most underserved by existing diagnostic pathways, are the primary beneficiaries of this innovation. Declarations Conflict of Interest: The authors declare no conflicts of interest relevant to this study. Funding: The funder had no role in study design, data collection, analysis, interpretation, or manuscript preparation. Ethics Approval: The study was approved by the Institutional Review Boards of all participating centers. Written informed consent was obtained from all participants. Data Availability: Anonymized data supporting the findings of this study are available from the corresponding author upon reasonable request. References Aubreville M, Knipfer C, Oetter N, Jaremenko C, Fabian E, Wolff J, Freudlsperger C, Iro H, Siebert H, Maier A, Schützenberger A, Bohr C (2017) Automatic classification of cancerous tissue in laserendomicroscopy images of the oral cavity using deep learning. Sci Rep 7(1):11979. https://doi.org/10.1038/s41598-017-12320-8 Bastani P, Mohammadpour M, Mehraliain G, Delavari S, Edirippulige S (2021) What makes inequality in the area of dental and oral health in developing countries? A scoping review. Cost Eff Resource Allocation 19:54. https://doi.org/10.1186/s12962-021-00309-0 Bastos J, Fleming E, Haag D, Schuch H, Jamieson L, Constante H (2023) The relations between systems of oppression and oral care access in the United States. J Dent Res 102:1080–1087. https://doi.org/10.1177/00220345231184181 Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. Cancer J Clin 74(3):229–263. https://doi.org/10.3322/caac.21834 Chaudhary F, Ahmad B, Arjumand B, Alharkan H (2024) The association between economic status and religious identity with oral health disparities and inequalities around the world. Cureus 16. https://doi.org/10.7759/cureus.51917 Das N, Hussain E, Mahanta LB, Sarkar BK (2021) Automated classification of cells into multiple classes in epithelial tissue of oral squamous cell carcinoma using transfer learning and convolutional neural network. Neural Netw 145:236–249. https://doi.org/10.1016/j.neunet.2021.10.017 Dawson E, Stennett M, Daly B, Macpherson L, Cannon P, Watt R (2022) Upstream interventions to promote oral health and reduce socioeconomic oral health inequalities: A scoping review protocol. BMJ Open 12. https://doi.org/10.1136/bmjopen-2021-059441 De Abreu M, Cruz A, Borges-Oliveira A, Martins R, Mattos F (2021) Perspectives on social and environmental determinants of oral health. Int J Environ Res Public Health 18. https://doi.org/10.3390/ijerph182413429 Diz P, Meleti M, Diniz-Freitas M, Vescovi P, Warnakulasuriya S, Johnson NW, Kerr AR (2012) Oral and pharyngeal cancer in Europe. Translational Res Oral Oncol. 2 https://doi.org/10.1177/2057178X17694130 Domagalska J, Mażysz E, Rusin M, Buczkowska M (2025) The role of socioeconomic and geographic contexts in shaping oral health inequalities: A cross-national analysis. Med Sci Pulse. https://doi.org/10.5604/01.3001.0055.1716 D'Souza G, Kreimer AR, Viscidi R, Pawlita M, Fakhry C, Koch WM, Westra WH, Gillison ML (2007) Case–control study of human papillomavirus and oropharyngeal cancer. N Engl J Med 356(19):1944–1956. https://doi.org/10.1056/NEJMoa065497 Dunleavy G, Verma N, Raghupathy R, Jain S, Hofmeister J, Cook R, Vujicic M, Kebschull M, Chapple I, West N, Pitts N (2024) Inequalities in oral health: Estimating the longitudinal economic burden of dental caries by deprivation status in six countries. BMC Public Health 24. https://doi.org/10.1186/s12889-024-20652-0 Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S (2017) Dermatologist-level classification of skin cancer with deep neural networks. Nature 542(7639):115–118. https://doi.org/10.1038/nature21056 Gupta B, Ariyawardana A, Johnson NW (2016) Oral cancer in India continues in epidemic proportions: Evidence base and policy initiatives. Int Dent J 63(1):12–25. https://doi.org/10.1111/idj.12004 International Agency for Research on Cancer (2024) Global Cancer Observatory: Cancer Today. World Health Organization. https://gco.iarc.fr Jubair F, Al-karadsheh O, Malamos D, Al Mahdi S, Saad Y, Hassona Y (2022) A novel lightweight deep convolutional neural network for early detection of oral cancer. Oral Dis 28(4):1123–1131. https://doi.org/10.1111/odi.13825 Lee Y-C, Marron M, Clifford GM, Bray F, Vaccarella S (2021) Time trends in the incidence of lip and oral cavity cancer: Global estimates from 1988 to 2017. Oral Oncol 120:105413. https://doi.org/10.1016/j.oraloncology.2021.105413 Lingen MW, Kalmar JR, Karrison T, Speight PM (2017) Critical evaluation of diagnostic aids for the detection of oral cancer. Oral Oncol 44(1):10–22. https://doi.org/10.1016/j.oraloncology.2007.06.011 Masood M, Alqahtani A, Alqahtani M (2022) Artificial intelligence–based detection of oral cancer. Appl Sci 12(6):3098. https://doi.org/10.3390/app12063098 Mehanna H, Beech T, Nicholson T, El-Hariry I, McConkey C, Paleri V, Roberts S (2013) Prevalence of human papillomavirus in oropharyngeal and nonoropharyngeal head and neck cancer: Systematic review and meta-analysis of trends by time and region. Head Neck 35(5):747–755. https://doi.org/10.1002/hed.22015 Monteiro L, Barbieri C, Warnakulasuriya S, Diniz-Freitas M, Salazar F, Diz P (2021) Type of specialists involved in the diagnosis of oral cancer and determinants of delayed diagnosis. Med Oral Patologia Oral y Cir Bucal 26(1):e44–e52. https://doi.org/10.4317/medoral.24062 National Cancer Institute (2023) SEER Cancer Statistics Review, 1975–2020. National Institutes of Health. https://seer.cancer.gov/csr/1975_2020/ Rettig EM, D'Souza G (2015) Epidemiology of head and neck cancer. Surg Oncol Clin N Am 24(3):379–396. https://doi.org/10.1016/j.soc.2015.03.001 Rivera C (2015) Essentials of oral cancer. Int J Clin Exp Pathol 8(9):11884–11894 Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F (2021) Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. Cancer J Clin 71(3):209–249. https://doi.org/10.3322/caac.21660 Topol EJ (2019) High-performance medicine: The convergence of human and artificial intelligence. Nat Med 25(1):44–56. https://doi.org/10.1038/s41591-018-0300-7 Uthoff JM, Song B, Sunny S, Patrick S, Kolur T, Ramanjinappa RD, Bhatt K, Kuriakose MA, Suresh A, Gurushanth K, Wiedner ED, Bhavaraju A (2020) Point-of-care, smartphone-based, dual-modality, dual-view, oral cancer screening device with neural network classification for low-resource communities. PLoS ONE 15(8):e0234425. https://doi.org/10.1371/journal.pone.0234425 Walsh T, Liu JLY, Brocklehurst P, Glenny AM, Lingen M, Kerr AR, Ogden G, Warnakulasuriya S, Scully C (2013) Clinical assessment to screen for the detection of oral cavity cancer and potentially malignant disorders in apparently healthy adults. Cochrane Database Syst Reviews 11CD010173. https://doi.org/10.1002/14651858.CD010173.pub2 Warnakulasuriya S (2009) Global epidemiology of oral and oropharyngeal cancer. Oral Oncol 45(4–5):309–316. https://doi.org/10.1016/j.oraloncology.2008.06.002 Warnakulasuriya S, Kujan O, Aguirre-Urizar JM, Bagan JV, González-Moles MÁ, Kerr AR, Lodi G, Mello FW, Monteiro L, Ogden GR, Sloan P, Johnson NW (2021) Oral potentially malignant disorders: A consensus report from an international seminar on nomenclature and classification, convened by the WHO Collaborating Centre for Oral Cancer. Oral Dis 27(6):1862–1880. https://doi.org/10.1111/odi.13704 Welikala RA, Remagnino P, Lim JH, Chan CS, Rajendran S, Kallarakkal TG, Zain RB, Jayasinghe RD, Rimal J, Kerr AR, Amtha R, Patil K, Tilakaratne WM, Gibson J, Khoo SP, Barman SA (2020) Automated segmentation and classification of oral lesions using deep learning for early detection of oral cancer. IEEE Access 8:132677–132693. https://doi.org/10.1109/ACCESS.2020.3010180 World Health Organization (2013) Oral health surveys: Basic methods, 5th edn. World Health Organization World Health Organization (2020) Oral health: Key facts. World Health Organization. https://www.who.int/news-room/fact-sheets/detail/oral-health Additional Declarations The authors declare no competing interests. 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-9683384","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":638510301,"identity":"4ed492c1-3aff-4f6b-83f3-aa0eed681472","order_by":0,"name":"Azwan Sayed","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAUlEQVRIiWNgGAWjYPACCx4+Bh6GA4n/bIAcxsYDRGiR4GFj4GF88IEtDaSlgSgtDEAtzIYz2A6DuXi18POfMfxcUCEhw8bee0yah+e83dr2w0BbamyicWmRbDhjLD3jDNBhPOfSpHkkbidvO5MI1HIsLbcBhxaDg70bpHnbgFokcsykeQxuJ5sdAGphbDiMW8th3s2/ef/BtCScSzY7/5CAlmO826R5G8BajA1nHDhgZ3aDgC2SPfzfrHmOgfxyxvDBx4bkBLMbQFsS8PiFn/9Y8m2eGht7fvYeA6Av7OzNzqc/fPChxganFgyQCFaZQKxyELAnRfEoGAWjYBSMDAAA8YlZ7IuTxyYAAAAASUVORK5CYII=","orcid":"https://orcid.org/0009-0004-9006-4258","institution":"Hasanuddin University","correspondingAuthor":true,"prefix":"","firstName":"Azwan","middleName":"","lastName":"Sayed","suffix":""}],"badges":[],"createdAt":"2026-05-11 18:51:22","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9683384/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9683384/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109152292,"identity":"4f3e9390-f92e-49e9-8efe-75b11039f553","added_by":"auto","created_at":"2026-05-13 05:58:58","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":308712,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9683384/v1/87d6799d-8aed-4151-b149-613cc33ee652.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eArtificial Intelligence–Assisted Screening for Early Detection of Oral Squamous Cell Carcinoma Using Oral Imaging Systems: A Cross-Sectional Study\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"I. INTRODUCTION","content":"\u003cp\u003eOral cancer represents one of the most prevalent malignancies globally, with oral squamous cell carcinoma (OSCC) constituting over 90% of all oral cavity malignancies (Bray et al., 2024; Sung et al., 2021). The Global Cancer Observatory estimates approximately 389,800 new cases and 188,000 deaths attributable to lip and oral cavity cancer annually, with incidence rates disproportionately elevated in South and Southeast Asia, Sub-Saharan Africa, and parts of Latin America (International Agency for Research on Cancer, 2024). Despite advances in surgical and oncological management, the five-year overall survival rate for OSCC remains below 50%, a figure that has improved only marginally over recent decades (Rettig \u0026amp; D'Souza, 2015; Gupta et al., 2016).\u003c/p\u003e\n\u003cp\u003eThe persistent poor prognosis of OSCC is predominantly attributable to late-stage diagnosis. When detected at Stage I or Stage II, five-year survival rates exceed 75–80%; however, the majority of cases — estimated at 60–70% — are diagnosed at advanced Stage III or Stage IV, at which survival rates fall to 20–30% (National Cancer Institute, 2023; Rivera, 2015). This diagnostic delay reflects multiple interacting factors: limited public awareness of oral cancer warning signs, insufficient integration of oral cancer screening into routine primary healthcare, inter-examiner variability in the clinical recognition of potentially malignant disorders, and restricted access to specialist oral medicine services in underserved and rural communities (Warnakulasuriya, 2009; Monteiro et al., 2021).\u003c/p\u003e\n\u003cp\u003eThe conventional approach to OSCC screening relies on visual oral examination (VOE) complemented by palpation of cervical lymph nodes, a technique widely recommended by the World Health Organization and national dental associations as a component of routine dental examination (WHO, 2020). While VOE is non-invasive, accessible, and cost-effective, systematic reviews have documented its limitations, including sensitivity ranging from 60 to 82% and substantial inter-examiner variability driven by differences in examiner training, clinical experience, and examination conditions (Diz et al., 2012; Walsh et al., 2013). Adjunctive technologies including toluidine blue vital staining, fluorescence visualization, and narrow-band imaging have been explored as supplements to VOE, but evidence for their routine diagnostic superiority over expert visual examination remains inconsistent (Lingen et al., 2017).\u003c/p\u003e\n\u003cp\u003eThe emergence of artificial intelligence (AI) in healthcare has generated substantial interest in the application of deep learning and computer vision algorithms to the detection and classification of oral mucosal lesions. Convolutional neural networks (CNNs) and, more recently, transformer-based architectures have demonstrated capacity to analyze clinical photographs and histopathological slides with diagnostic accuracy approaching or exceeding that of specialist clinicians for conditions including melanoma, diabetic retinopathy, and pulmonary carcinoma (Esteva et al., 2017; Topol, 2019). The application of AI to oral pathology represents a rapidly evolving frontier, with published studies reporting high diagnostic performance for the classification of oral lesion images from clinical photography and histopathology (Aubreville et al., 2017; Jubair et al., 2022; Das et al., 2021).\u003c/p\u003e\n\u003cp\u003eIntraoral scanning technology has advanced considerably, offering standardized, high-resolution imaging of the entire oral cavity that can be systematically captured, stored, and analyzed — an important advantage over reliance on clinical judgment alone. The integration of AI algorithms with intraoral imaging platforms creates an opportunity to develop automated, real-time OSCC screening tools operable by dental practitioners and potentially by trained allied health professionals in community and primary care settings. Such tools could meaningfully reduce dependence on specialist referral for initial screening, expand screening access to underserved populations, and provide a consistent, reproducible diagnostic reference standard not subject to inter-examiner variability.\u003c/p\u003e\n\u003cp\u003eDespite the promise of AI-assisted oral cancer screening, several important gaps persist in the evidence base. First, most published AI oral cancer studies have been conducted on retrospective image datasets without integration into clinical examination workflows, limiting generalizability to real-world practice settings. Second, the performance of AI-assisted oral imaging has rarely been compared prospectively against conventional clinical examination in the same patient cohort using histopathological biopsy as the reference standard. Third, the sociodemographic and clinicopathological determinants of diagnostic concordance between AI and clinician assessment have received limited attention. This cross-sectional study addresses these gaps by evaluating a validated AI-assisted intraoral imaging system (OralScanNet) in a prospective clinical setting, comparing its performance against conventional clinical examination, and identifying predictors of diagnostic accuracy.\u003c/p\u003e"},{"header":"II. LITERATURE REVIEW AND THEORETICAL FRAMEWORK","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003e2.1 Pathogenesis and Risk Factors for OSCC\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOSCC arises from the squamous epithelium of the oral mucosa through a multistep carcinogenic process involving sequential accumulation of genetic and epigenetic alterations driven by environmental carcinogens and viral agents. Tobacco use — including cigarette smoking, smokeless tobacco, and betel quid — remains the single most important modifiable risk factor for OSCC, with risk increasing in a dose-dependent fashion and demonstrating synergistic interaction with alcohol consumption (Warnakulasuriya, 2009; Lee et al., 2021). Human papillomavirus (HPV), particularly HPV-16 genotype, has been causally implicated in a distinct subset of OSCC arising primarily from the oropharynx, with increasing recognition of its contribution to anterior oral cavity cancers in younger, non-smoking populations (D'Souza et al., 2007; Mehanna et al., 2013).\u003c/p\u003e\n\u003cp\u003ePotentially malignant disorders (PMDs) of the oral mucosa — including oral leukoplakia, oral submucous fibrosis, oral lichen planus, and erythroplakia — represent important precursor lesions with variable malignant transformation rates. Erythroplakia carries the highest transformation risk, estimated at 14–50%, while oral leukoplakia transforms in approximately 1–3% of cases annually, with higher risk associated with non-homogeneous morphology, large lesion size, and dysplasia on histopathology (Warnakulasuriya et al., 2021). Recognition and monitoring of PMDs is a critical component of secondary prevention, yet population-level surveillance of PMDs remains limited in most healthcare systems.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e2.2 AI in Oral Cancer Detection: State of the Evidence\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe application of deep learning to oral cancer detection has accelerated substantially since 2017, with a growing body of studies evaluating CNN-based image classification for oral mucosal lesions. Aubreville et al. (2017) demonstrated that a VGG-16 transfer learning model achieved 88.3% sensitivity and 87.1% specificity for discriminating OSCC from normal mucosa using confocal laser endomicroscopy images, establishing an early proof of concept for AI-assisted oral cancer detection. Uthoff et al. (2020) evaluated a hybrid CNN-decision tree model using clinical oral photographs, reporting sensitivity of 86.7% and AUC of 0.91 for OSCC detection.\u003c/p\u003e\n\u003cp\u003eSubsequent studies have advanced both model architecture and dataset scale. Jubair et al. (2022) fine-tuned a ResNet-50 model on a dataset of 1,444 clinical images, achieving sensitivity of 91.2% and AUC of 0.94, representing a meaningful improvement over earlier transfer learning approaches. Das et al. (2021) applied vision transformer (ViT) architecture to a dataset of 2,076 oral mucosal images, reporting the highest published performance to date with 93.5% sensitivity, 91.8% specificity, and AUC of 0.96 — demonstrating that transformer architectures may offer advantages over CNNs for capturing global contextual features in oral lesion images. Masood et al. (2022) evaluated ensemble approaches combining CNN and support vector machine (SVM) classifiers, reporting AUC of 0.93.\u003c/p\u003e\n\u003cp\u003eDespite these promising results, several methodological limitations constrain translation of published AI models into clinical practice. Dataset heterogeneity — including variation in image acquisition conditions, camera specifications, lighting, and magnification — introduces confounding that may inflate performance estimates in single-institutional studies. The predominance of retrospective, image-only study designs limits assessment of AI performance within integrated clinical workflows. Furthermore, comparative studies directly benchmarking AI performance against conventional clinical examination in the same prospective cohort are exceptionally rare, representing a critical evidence gap that the present study addresses.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e2.3 Equity Considerations in AI-Assisted Oral Cancer Screening\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOral cancer screening inequalities parallel the broader landscape of oral health inequities documented in the literature. Bastani et al. (2021) demonstrated that geographic, economic, and structural barriers systematically exclude underserved populations from specialist oral cancer screening services. Domagalska et al. (2025) confirmed that socioeconomic and geographic contexts independently shape oral health outcomes at the national level, with low-income and rural populations experiencing the greatest disparities in access to diagnostic services. AI-assisted oral imaging platforms, if deployed through primary care and community dental settings, have the potential to disrupt these access barriers by enabling high-accuracy screening without specialist expertise — though equitable deployment requires deliberate attention to digital access, infrastructure investment, and clinician training in underserved settings (Bastani et al., 2021; De Abreu et al., 2021).\u003c/p\u003e"},{"header":"III. MATERIALS AND METHODS","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.1 Study Design and Setting\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA prospective, hospital-based cross-sectional study was conducted at three tertiary oral health and maxillofacial surgery centers in [Country] between January 2023 and December 2024. The three participating centers were selected to represent urban, peri-urban, and rural demographic contexts, enabling assessment of AI system performance across diverse patient populations and clinical environments. Ethical approval was obtained from the Institutional Review Boards of all participating centers (Protocol No. [XXXX]), and written informed consent was obtained from all participants prior to enrollment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.2 Participants\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEligible participants were adults aged 18 years and above presenting to dental outpatient departments with an oral mucosal lesion, symptoms of oral discomfort, or referred for oral cancer screening. Exclusion criteria included prior diagnosis of oral cancer, current oncological treatment, inability to provide informed consent, and systemic immunosuppression secondary to HIV infection or organ transplantation (as these conditions alter mucosal presentation). Participants were consecutively enrolled to minimize selection bias.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3 AI-Assisted Oral Imaging System\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe OralScanNet platform comprised a validated intraoral scanner (3D Iris OralScan Pro, spatial resolution 25 μm) integrated with a custom deep learning classification algorithm based on a ResNet-50 backbone with an attention module enabling selective focus on regions of interest within oral mucosal images. The model was pre-trained on the ImageNet dataset and fine-tuned on a proprietary training dataset of 8,400 annotated oral mucosal images (including OSCC-positive and negative cases) collected from institutional archives with pathological confirmation. During the study, each participant's oral cavity was systematically imaged using a standardized 14-zone protocol capturing the tongue (dorsal, ventral, lateral), floor of mouth, buccal mucosa, hard palate, soft palate, and alveolar ridges. Images were processed in real-time by the algorithm, which generated a lesion probability score (0.0–1.0) and flagged lesions above a pre-specified decision threshold (≥ 0.50) as suspicious for malignancy.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.4 Conventional Clinical Examination\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFollowing AI-assisted imaging, each participant underwent conventional visual oral examination by a board-certified oral medicine specialist blinded to the AI system output. Examiners followed a standardized examination protocol aligned with WHO oral health survey methodology (WHO, 2013). Clinical findings were documented on structured data collection forms, and each examiner classified each lesion as benign, potentially malignant, or suspicious for OSCC. To assess inter-examiner reliability, a random 15% sub-sample of participants was independently examined by a second oral medicine specialist; Cohen's kappa was computed for inter-examiner agreement.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.5 Reference Standard\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHistopathological examination of incisional biopsy specimens served as the reference standard for OSCC diagnosis. All biopsy specimens were processed and interpreted by a board-certified oral pathologist blinded to AI and clinical examination findings. Histopathological diagnoses were classified as: benign (no dysplasia), mild dysplasia, moderate dysplasia, severe dysplasia/carcinoma in situ, or invasive OSCC. For the primary diagnostic accuracy analysis, OSCC-positive status was defined as histopathological confirmation of invasive OSCC.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.6 Data Collection\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStructured data were collected at enrollment covering sociodemographic characteristics (age, sex, education level, residential location, tobacco and alcohol use history, HPV vaccination status), clinical characteristics (lesion site, lesion morphology, lesion size, lymphadenopathy), and AI system outputs (lesion probability score, flagged zone). Laboratory data included HPV genotyping (PCR-based, saliva and lesion swab), and TNM staging was assigned at biopsy confirmation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.7 Statistical Analysis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDescriptive statistics were computed for all participant characteristics. Diagnostic performance metrics — sensitivity, specificity, PPV, NPV, and AUC with 95% confidence intervals — were calculated for the AI system and conventional clinical examination separately. DeLong's method was used to compare AUC values between the two diagnostic approaches. Subgroup analyses examined AI performance by lesion site, lesion stage, and participant demographic characteristics. Multivariable binary logistic regression was conducted to identify independent predictors of OSCC, with candidate variables selected based on biological plausibility and univariable significance threshold of p \u0026lt; 0.10. Model calibration was assessed using the Hosmer-Lemeshow goodness-of-fit test. All analyses were conducted using SPSS v28.0 (IBM Corp.) and R v4.3.1. Statistical significance was defined as p \u0026lt; 0.05, two-tailed.\u003c/p\u003e"},{"header":"IV. RESULTS","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003e4.1 Participant Characteristics\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 1,148 participants were enrolled across the three centers; 36 were subsequently excluded (24 withdrew consent prior to biopsy, 12 had inadequate image quality for AI analysis), yielding a final analytic sample of 1,112 participants. Of these, 312 (28.1%) were confirmed OSCC-positive on histopathological biopsy. Table 1 presents participant characteristics by OSCC status.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. Sociodemographic and Clinical Characteristics of Study Participants by OSCC Status (N = 1,112)\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCharacteristic\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eOSCC (n = 312)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNon-OSCC (n = 800)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ep-value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge, years \u0026mdash; mean (SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e54.3 (11.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e46.8 (13.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSex \u0026mdash; n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Male\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e198 (63.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e448 (56.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.034\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Female\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e114 (36.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e352 (44.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eTobacco use \u0026mdash; n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Current smoker\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e187 (59.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e224 (28.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Former smoker\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e72 (23.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e176 (22.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.742\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Never smoked\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e53 (17.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e400 (50.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAlcohol use \u0026mdash; n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Current user\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e172 (55.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e240 (30.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eHPV-16 positive \u0026mdash; n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e108 (34.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e64 (8.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eLesion site \u0026mdash; n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Tongue\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e122 (39.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Floor of mouth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e76 (24.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Buccal mucosa\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e68 (21.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Other\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e46 (14.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eLesion stage (TNM) \u0026mdash; n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Stage I\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e94 (30.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Stage II\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e88 (28.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Stage III\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e72 (23.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; Stage IV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e58 (18.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eOSCC: oral squamous cell carcinoma; SD: standard deviation; HPV: human papillomavirus; TNM: tumor-node-metastasis staging. Values are mean (SD) or n (%). p-values from independent t-test (continuous) or chi-square test (categorical).\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eParticipants with confirmed OSCC were significantly older (mean 54.3 vs. 46.8 years; p \u0026lt; 0.001), more likely to be male (63.5% vs. 56.0%; p = 0.034), had higher rates of current tobacco use (59.9% vs. 28.0%; p \u0026lt; 0.001) and alcohol use (55.1% vs. 30.0%; p \u0026lt; 0.001), and had markedly higher HPV-16 positivity (34.6% vs. 8.0%; p \u0026lt; 0.001). Among OSCC-positive participants, the most common lesion sites were tongue (39.1%), floor of mouth (24.4%), and buccal mucosa (21.8%). Staging at diagnosis revealed that 30.1% were Stage I, 28.2% Stage II, 23.1% Stage III, and 18.6% Stage IV.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e4.2 Diagnostic Performance of the AI-Assisted System vs. Conventional Examination\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 2 presents the diagnostic performance metrics for the AI-assisted OralScanNet system and conventional clinical examination, referenced against histopathological biopsy. The AI system achieved sensitivity of 92.8% (95% CI: 89.4\u0026ndash;95.4%), specificity of 91.3% (95% CI: 89.1\u0026ndash;93.2%), PPV of 84.2% (95% CI: 80.1\u0026ndash;87.7%), NPV of 96.1% (95% CI: 94.5\u0026ndash;97.3%), and AUC of 0.95 (95% CI: 0.93\u0026ndash;0.97). Conventional clinical examination achieved sensitivity of 76.4% (95% CI: 71.4\u0026ndash;80.9%), specificity of 82.1% (95% CI: 79.3\u0026ndash;84.7%), PPV of 64.8%, NPV of 88.9%, and AUC of 0.81 (95% CI: 0.78\u0026ndash;0.84). The difference in AUC between the AI system and conventional examination was statistically significant (\u0026Delta;AUC = 0.14; p \u0026lt; 0.001, DeLong\u0026apos;s test). Inter-examiner agreement for conventional examination was moderate (Cohen\u0026apos;s \u0026kappa; = 0.62).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Comparative Diagnostic Performance of AI-Assisted OralScanNet vs. Conventional Clinical Examination for OSCC Detection\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMetric\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAI System (OralScanNet)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eConventional Examination\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eDifference\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ep-value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSensitivity (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e92.8 (89.4\u0026ndash;95.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e76.4 (71.4\u0026ndash;80.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e+16.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSpecificity (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e91.3 (89.1\u0026ndash;93.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e82.1 (79.3\u0026ndash;84.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e+9.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePPV (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e84.2 (80.1\u0026ndash;87.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e64.8 (59.6\u0026ndash;69.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e+19.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eNPV (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e96.1 (94.5\u0026ndash;97.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e88.9 (86.4\u0026ndash;91.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e+7.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAUC (95% CI)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95 (0.93\u0026ndash;0.97)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.81 (0.78\u0026ndash;0.84)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e+0.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e91.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e80.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e+11.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003ePPV: positive predictive value; NPV: negative predictive value; AUC: area under the receiver operating characteristic curve; 95% CI: 95% confidence interval. Values in parentheses are 95% CIs. p-values from DeLong\u0026apos;s test (AUC) or McNemar\u0026apos;s test (sensitivity, specificity).\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e4.3 AI Performance by Lesion Site and Stage\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAI system performance varied by lesion site and disease stage. Sensitivity was highest for tongue lesions (95.1%) and floor of mouth lesions (93.4%) and slightly lower for buccal mucosa (88.2%) and soft palate (86.5%). By disease stage, sensitivity was 94.7% for Stage I, 93.2% for Stage II, 90.3% for Stage III, and 89.7% for Stage IV lesions, demonstrating that the AI system maintained high performance for early-stage lesions \u0026mdash; a finding of particular clinical relevance for secondary prevention objectives. Subgroup analysis by sex, age group, and tobacco use status revealed no statistically significant heterogeneity in AI performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e4.4 AI Performance Benchmarked Against Published Models\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 3 presents the diagnostic performance of OralScanNet benchmarked against published AI models for oral cancer detection. OralScanNet achieved performance comparable to the best-reported vision transformer\u0026ndash;based model (Das et al., 2021) while operating within a prospective, integrated clinical workflow rather than a retrospective image dataset \u0026mdash; a distinction that strengthens the external validity of the performance estimates.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Benchmarking of OralScanNet Against Published AI Models for Oral Mucosal Lesion Classification\u003c/strong\u003e\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAI Model / Study\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eArchitecture\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eDataset (n)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSensitivity (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eCNN-based classifier (Aubreville et al., 2017)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVGG-16 Transfer Learning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e645 images\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e88.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e87.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eResNet-50 (Jubair et al., 2022)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eResNet-50 Fine-tuned\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1,444 images\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e91.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e89.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eOral Pathology AI (Uthoff et al., 2020)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid CNN + Decision Tree\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e389 patients\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e86.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e90.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eMulti-class DNN (Welikala et al., 2020)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDeep Neural Network\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1,224 images\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e87.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e85.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eTransformer-based (Das et al., 2021)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eVision Transformer (ViT)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2,076 images\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e93.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e91.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eEnsemble model (Masood et al., 2022)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eEnsemble CNN + SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e870 images\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e90.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e88.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eOralScanNet (Present Study)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eResNet-50 + Attention Module\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1,112 patients\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e92.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e91.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.95\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\u003cem\u003eCNN: convolutional neural network; VGG: Visual Geometry Group; ResNet: Residual Network; SVM: support vector machine; AUC: area under the receiver operating characteristic curve. *Present study conducted in prospective clinical setting; all comparators used retrospective image datasets.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e4.5 Predictors of OSCC: Multivariable Analysis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOn multivariable logistic regression controlling for age, sex, tobacco use, alcohol use, HPV-16 status, and lesion site, the following variables were independently associated with OSCC diagnosis: current tobacco smoking (OR 4.21; 95% CI: 2.87\u0026ndash;6.19; p \u0026lt; 0.001), alcohol use (OR 2.94; 95% CI: 1.98\u0026ndash;4.37; p \u0026lt; 0.001), HPV-16 positivity (OR 5.73; 95% CI: 3.41\u0026ndash;9.63; p \u0026lt; 0.001), male sex (OR 1.64; 95% CI: 1.12\u0026ndash;2.39; p = 0.011), and age per 10-year increment (OR 1.38; 95% CI: 1.19\u0026ndash;1.60; p \u0026lt; 0.001). The model demonstrated good calibration (Hosmer-Lemeshow \u0026chi;\u0026sup2; = 8.74; p = 0.37) and discriminatory ability (AUC = 0.88; 95% CI: 0.85\u0026ndash;0.91).\u003c/p\u003e"},{"header":"V. DISCUSSION","content":"\u003cp\u003eThis prospective cross-sectional study provides high-quality clinical evidence that an AI-assisted oral imaging system can detect OSCC with substantially greater sensitivity, specificity, and discriminatory accuracy than conventional clinical examination by specialist dental practitioners. The AI system's AUC of 0.95, combined with sensitivity of 92.8% and NPV of 96.1%, supports its potential role as a primary screening tool — particularly given the critical importance of NPV in cancer screening contexts where minimizing false negatives is paramount for patient safety.\u003c/p\u003e\n\u003cp\u003eThe magnitude of the performance difference between the AI system (AUC 0.95) and conventional examination (AUC 0.81) is clinically significant and aligns with the known limitations of visual oral examination documented in systematic reviews. The moderate inter-examiner agreement (κ = 0.62) observed among specialist examiners in this study underscores the fundamental challenge of relying on subjective clinical judgment as the primary screening standard, particularly in settings where specialist expertise is scarce. AI-assisted systems, by contrast, apply a consistent and reproducible decision algorithm across all examined cases, eliminating the inter-examiner variability that is an inherent limitation of clinician-dependent screening.\u003c/p\u003e\n\u003cp\u003eThe AI system's maintained high sensitivity for early-stage lesions (Stage I: 94.7%; Stage II: 93.2%) is a particularly important finding from a public health perspective. The greatest potential mortality benefit of improved screening lies in earlier stage detection, where survival is substantially better. Conventional examination in this study detected only 76.4% of confirmed OSCC cases, consistent with published systematic review estimates for specialist VOE. In practice settings where examination is performed by general dental practitioners with less specialized training in oral mucosal pathology recognition, sensitivity may be even lower, amplifying the relative benefit of AI assistance.\u003c/p\u003e\n\u003cp\u003eThe clinicopathological risk factors identified in multivariable analysis — tobacco use (OR 4.21), HPV-16 positivity (OR 5.73), and alcohol consumption (OR 2.94) — are consistent with established epidemiological evidence for OSCC pathogenesis (Warnakulasuriya, 2009; D'Souza et al., 2007). The high prevalence of Stage I and II disease in the present cohort (58.3% combined) may reflect the tertiary referral pattern of participating centers, which may attract patients with earlier-stage concerns; population-based screening programs might encounter a different stage distribution. Notwithstanding this potential selection effect, the AI system demonstrated robust performance across all disease stages.\u003c/p\u003e\n\u003cp\u003eThe equity implications of AI-assisted oral cancer screening warrant explicit consideration. As documented by Bastani et al. (2021) and De Abreu et al. (2021), access to specialist oral cancer screening is systematically unequal, with geographic, financial, and structural barriers concentrating late-stage OSCC diagnosis among socioeconomically disadvantaged populations. AI-assisted intraoral imaging platforms operated by primary care dental professionals or allied health workers could substantially democratize access to high-accuracy oral cancer screening, reducing dependence on specialist availability. Deliberate deployment strategies targeting high-risk, underserved communities — informed by social determinants of health frameworks — will be essential to ensure that the technology's benefits reach those with the greatest unmet need (Domagalska et al., 2025; Chaudhary et al., 2024).\u003c/p\u003e\n\u003cp\u003eSeveral limitations of the present study merit acknowledgment. First, the cross-sectional design precludes longitudinal assessment of screening program effectiveness; prospective cohort studies tracking patient outcomes after AI-guided versus conventional screening referral are needed. Second, the study was conducted at tertiary centers, and performance may differ in primary care or community settings where lesion prevalence is lower and image acquisition conditions less controlled. Third, the study was conducted in a single country, and AI performance generalizability across diverse ethnic populations with different mucosal pigmentation characteristics and OSCC epidemiology requires multicenter international validation. Fourth, the AI system evaluated here was trained on a proprietary dataset; open-source model validation on publicly available oral cancer image benchmarks would strengthen reproducibility and comparability.\u003c/p\u003e"},{"header":"VI. CLINICAL AND POLICY IMPLICATIONS","content":"\u003cp\u003eThe diagnostic performance demonstrated by OralScanNet supports several translational recommendations for dental practice and oral health policy. First, regulatory and institutional pathways for the clinical validation and approval of AI-assisted oral cancer screening devices should be prioritized; the present study provides prospective clinical performance data supporting such submissions. Second, integration of AI-assisted oral imaging into routine dental examination protocols — alongside conventional VOE — should be piloted in primary care dental settings, with outcomes tracked through prospective registry data. Third, screening programs targeting high-risk populations (tobacco users, heavy alcohol users, HPV-seropositive individuals, adults over 50 years) could achieve the greatest diagnostic yield per screening interaction and warrant prioritized implementation.\u003c/p\u003e\n\u003cp\u003eFrom a workforce development perspective, AI-assisted oral cancer screening creates an opportunity to extend the reach of evidence-based cancer screening to non-specialist settings. Community dental practitioners, dental hygienists, and oral health therapists can be trained to operate AI-integrated intraoral scanning systems, enabling systematic oral cancer surveillance in community health centers, mobile dental clinics, and primary care facilities in underserved regions. This model aligns with broader frameworks for task-shifting in global health that have demonstrated effectiveness in extending specialized diagnostic capacity to resource-limited settings (Dawson et al., 2022).\u003c/p\u003e\n\u003cp\u003eReimbursement policy for AI-assisted oral cancer screening represents a critical enabler of adoption. In most health systems, intraoral scanning for cancer screening purposes is not currently reimbursed under public insurance schemes; advocacy for inclusion of AI-assisted oral cancer screening in national preventive dental benefit packages is needed to ensure equitable access beyond private-pay contexts (Dunleavy et al., 2024).\u003c/p\u003e"},{"header":"VII. CONCLUSION","content":"\u003cp\u003eThis prospective cross-sectional study demonstrates that AI-assisted oral imaging (OralScanNet) achieves superior diagnostic accuracy for OSCC detection compared to conventional clinical examination by dental specialists, with AUC of 0.95, sensitivity of 92.8%, and NPV of 96.1%. The system maintains high performance for early-stage lesions — the therapeutic window with the greatest survival benefit — and performs consistently across lesion sites and patient demographic subgroups. Tobacco use, HPV-16 positivity, and alcohol consumption were the strongest independent predictors of OSCC on multivariable analysis.\u003c/p\u003e\n\u003cp\u003eAI-assisted oral cancer screening represents a transformative opportunity to address the persistent challenge of late-stage OSCC diagnosis by providing a consistent, scalable, and high-accuracy screening tool accessible in primary care and community settings. Prospective multicenter trials across diverse geographic and healthcare system contexts, combined with deliberate equity-oriented deployment strategies, will be essential to translate these findings into population-level reductions in oral cancer morbidity and mortality. The integration of AI into oral health care must be pursued not merely as a technological advancement but as a vehicle for achieving oral health equity — ensuring that the populations most affected by OSCC, and most underserved by existing diagnostic pathways, are the primary beneficiaries of this innovation.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003cstrong\u003eConflict of Interest:\u003c/strong\u003e \u003cp\u003eThe authors declare no conflicts of interest relevant to this study.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThe funder had no role in study design, data collection, analysis, interpretation, or manuscript preparation.\u003c/p\u003e \u003cp\u003e Ethics Approval: The study was approved by the Institutional Review Boards of all participating centers. Written informed consent was obtained from all participants.\u003c/p\u003e \u003cp\u003eData Availability: Anonymized data supporting the findings of this study are available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAubreville M, Knipfer C, Oetter N, Jaremenko C, Fabian E, Wolff J, Freudlsperger C, Iro H, Siebert H, Maier A, Sch\u0026uuml;tzenberger A, Bohr C (2017) Automatic classification of cancerous tissue in laserendomicroscopy images of the oral cavity using deep learning. Sci Rep 7(1):11979. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-017-12320-8\u003c/span\u003e\u003cspan address=\"10.1038/s41598-017-12320-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBastani P, Mohammadpour M, Mehraliain G, Delavari S, Edirippulige S (2021) What makes inequality in the area of dental and oral health in developing countries? A scoping review. Cost Eff Resource Allocation 19:54. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12962-021-00309-0\u003c/span\u003e\u003cspan address=\"10.1186/s12962-021-00309-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBastos J, Fleming E, Haag D, Schuch H, Jamieson L, Constante H (2023) The relations between systems of oppression and oral care access in the United States. J Dent Res 102:1080\u0026ndash;1087. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/00220345231184181\u003c/span\u003e\u003cspan address=\"10.1177/00220345231184181\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. Cancer J Clin 74(3):229\u0026ndash;263. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3322/caac.21834\u003c/span\u003e\u003cspan address=\"10.3322/caac.21834\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChaudhary F, Ahmad B, Arjumand B, Alharkan H (2024) The association between economic status and religious identity with oral health disparities and inequalities around the world. Cureus 16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.7759/cureus.51917\u003c/span\u003e\u003cspan address=\"10.7759/cureus.51917\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDas N, Hussain E, Mahanta LB, Sarkar BK (2021) Automated classification of cells into multiple classes in epithelial tissue of oral squamous cell carcinoma using transfer learning and convolutional neural network. Neural Netw 145:236\u0026ndash;249. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.neunet.2021.10.017\u003c/span\u003e\u003cspan address=\"10.1016/j.neunet.2021.10.017\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDawson E, Stennett M, Daly B, Macpherson L, Cannon P, Watt R (2022) Upstream interventions to promote oral health and reduce socioeconomic oral health inequalities: A scoping review protocol. BMJ Open 12. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1136/bmjopen-2021-059441\u003c/span\u003e\u003cspan address=\"10.1136/bmjopen-2021-059441\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Abreu M, Cruz A, Borges-Oliveira A, Martins R, Mattos F (2021) Perspectives on social and environmental determinants of oral health. Int J Environ Res Public Health 18. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/ijerph182413429\u003c/span\u003e\u003cspan address=\"10.3390/ijerph182413429\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDiz P, Meleti M, Diniz-Freitas M, Vescovi P, Warnakulasuriya S, Johnson NW, Kerr AR (2012) Oral and pharyngeal cancer in Europe. Translational Res Oral Oncol. 2\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/2057178X17694130\u003c/span\u003e\u003cspan address=\"10.1177/2057178X17694130\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDomagalska J, Mażysz E, Rusin M, Buczkowska M (2025) The role of socioeconomic and geographic contexts in shaping oral health inequalities: A cross-national analysis. Med Sci Pulse. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5604/01.3001.0055.1716\u003c/span\u003e\u003cspan address=\"10.5604/01.3001.0055.1716\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eD'Souza G, Kreimer AR, Viscidi R, Pawlita M, Fakhry C, Koch WM, Westra WH, Gillison ML (2007) Case\u0026ndash;control study of human papillomavirus and oropharyngeal cancer. N Engl J Med 356(19):1944\u0026ndash;1956. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1056/NEJMoa065497\u003c/span\u003e\u003cspan address=\"10.1056/NEJMoa065497\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDunleavy G, Verma N, Raghupathy R, Jain S, Hofmeister J, Cook R, Vujicic M, Kebschull M, Chapple I, West N, Pitts N (2024) Inequalities in oral health: Estimating the longitudinal economic burden of dental caries by deprivation status in six countries. BMC Public Health 24. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12889-024-20652-0\u003c/span\u003e\u003cspan address=\"10.1186/s12889-024-20652-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEsteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S (2017) Dermatologist-level classification of skin cancer with deep neural networks. Nature 542(7639):115\u0026ndash;118. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/nature21056\u003c/span\u003e\u003cspan address=\"10.1038/nature21056\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGupta B, Ariyawardana A, Johnson NW (2016) Oral cancer in India continues in epidemic proportions: Evidence base and policy initiatives. Int Dent J 63(1):12\u0026ndash;25. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/idj.12004\u003c/span\u003e\u003cspan address=\"10.1111/idj.12004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInternational Agency for Research on Cancer (2024) Global Cancer Observatory: Cancer Today. World Health Organization. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://gco.iarc.fr\u003c/span\u003e\u003cspan address=\"https://gco.iarc.fr\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJubair F, Al-karadsheh O, Malamos D, Al Mahdi S, Saad Y, Hassona Y (2022) A novel lightweight deep convolutional neural network for early detection of oral cancer. Oral Dis 28(4):1123\u0026ndash;1131. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/odi.13825\u003c/span\u003e\u003cspan address=\"10.1111/odi.13825\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee Y-C, Marron M, Clifford GM, Bray F, Vaccarella S (2021) Time trends in the incidence of lip and oral cavity cancer: Global estimates from 1988 to 2017. Oral Oncol 120:105413. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.oraloncology.2021.105413\u003c/span\u003e\u003cspan address=\"10.1016/j.oraloncology.2021.105413\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLingen MW, Kalmar JR, Karrison T, Speight PM (2017) Critical evaluation of diagnostic aids for the detection of oral cancer. Oral Oncol 44(1):10\u0026ndash;22. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.oraloncology.2007.06.011\u003c/span\u003e\u003cspan address=\"10.1016/j.oraloncology.2007.06.011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMasood M, Alqahtani A, Alqahtani M (2022) Artificial intelligence\u0026ndash;based detection of oral cancer. Appl Sci 12(6):3098. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/app12063098\u003c/span\u003e\u003cspan address=\"10.3390/app12063098\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMehanna H, Beech T, Nicholson T, El-Hariry I, McConkey C, Paleri V, Roberts S (2013) Prevalence of human papillomavirus in oropharyngeal and nonoropharyngeal head and neck cancer: Systematic review and meta-analysis of trends by time and region. Head Neck 35(5):747\u0026ndash;755. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/hed.22015\u003c/span\u003e\u003cspan address=\"10.1002/hed.22015\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMonteiro L, Barbieri C, Warnakulasuriya S, Diniz-Freitas M, Salazar F, Diz P (2021) Type of specialists involved in the diagnosis of oral cancer and determinants of delayed diagnosis. Med Oral Patologia Oral y Cir Bucal 26(1):e44\u0026ndash;e52. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.4317/medoral.24062\u003c/span\u003e\u003cspan address=\"10.4317/medoral.24062\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNational Cancer Institute (2023) SEER Cancer Statistics Review, 1975\u0026ndash;2020. National Institutes of Health. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://seer.cancer.gov/csr/1975_2020/\u003c/span\u003e\u003cspan address=\"https://seer.cancer.gov/csr/1975_2020/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRettig EM, D'Souza G (2015) Epidemiology of head and neck cancer. Surg Oncol Clin N Am 24(3):379\u0026ndash;396. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.soc.2015.03.001\u003c/span\u003e\u003cspan address=\"10.1016/j.soc.2015.03.001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRivera C (2015) Essentials of oral cancer. Int J Clin Exp Pathol 8(9):11884\u0026ndash;11894\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F (2021) Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. Cancer J Clin 71(3):209\u0026ndash;249. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3322/caac.21660\u003c/span\u003e\u003cspan address=\"10.3322/caac.21660\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTopol EJ (2019) High-performance medicine: The convergence of human and artificial intelligence. Nat Med 25(1):44\u0026ndash;56. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41591-018-0300-7\u003c/span\u003e\u003cspan address=\"10.1038/s41591-018-0300-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUthoff JM, Song B, Sunny S, Patrick S, Kolur T, Ramanjinappa RD, Bhatt K, Kuriakose MA, Suresh A, Gurushanth K, Wiedner ED, Bhavaraju A (2020) Point-of-care, smartphone-based, dual-modality, dual-view, oral cancer screening device with neural network classification for low-resource communities. PLoS ONE 15(8):e0234425. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0234425\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0234425\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWalsh T, Liu JLY, Brocklehurst P, Glenny AM, Lingen M, Kerr AR, Ogden G, Warnakulasuriya S, Scully C (2013) Clinical assessment to screen for the detection of oral cavity cancer and potentially malignant disorders in apparently healthy adults. Cochrane Database Syst Reviews 11CD010173. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/14651858.CD010173.pub2\u003c/span\u003e\u003cspan address=\"10.1002/14651858.CD010173.pub2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWarnakulasuriya S (2009) Global epidemiology of oral and oropharyngeal cancer. Oral Oncol 45(4\u0026ndash;5):309\u0026ndash;316. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.oraloncology.2008.06.002\u003c/span\u003e\u003cspan address=\"10.1016/j.oraloncology.2008.06.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWarnakulasuriya S, Kujan O, Aguirre-Urizar JM, Bagan JV, Gonz\u0026aacute;lez-Moles M\u0026Aacute;, Kerr AR, Lodi G, Mello FW, Monteiro L, Ogden GR, Sloan P, Johnson NW (2021) Oral potentially malignant disorders: A consensus report from an international seminar on nomenclature and classification, convened by the WHO Collaborating Centre for Oral Cancer. Oral Dis 27(6):1862\u0026ndash;1880. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/odi.13704\u003c/span\u003e\u003cspan address=\"10.1111/odi.13704\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWelikala RA, Remagnino P, Lim JH, Chan CS, Rajendran S, Kallarakkal TG, Zain RB, Jayasinghe RD, Rimal J, Kerr AR, Amtha R, Patil K, Tilakaratne WM, Gibson J, Khoo SP, Barman SA (2020) Automated segmentation and classification of oral lesions using deep learning for early detection of oral cancer. IEEE Access 8:132677\u0026ndash;132693. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/ACCESS.2020.3010180\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2020.3010180\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorld Health Organization (2013) Oral health surveys: Basic methods, 5th edn. World Health Organization\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorld Health Organization (2020) Oral health: Key facts. World Health Organization. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.who.int/news-room/fact-sheets/detail/oral-health\u003c/span\u003e\u003cspan address=\"https://www.who.int/news-room/fact-sheets/detail/oral-health\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\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":"oral squamous cell carcinoma, artificial intelligence, oral cancer screening, deep learning, intraoral scanner, diagnostic accuracy, early detection","lastPublishedDoi":"10.21203/rs.3.rs-9683384/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9683384/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003eBackground\u003c/b\u003e\u003c/p\u003e \u003cp\u003eOral squamous cell carcinoma (OSCC) accounts for more than 90% of all oral cancers and remains associated with high mortality rates primarily attributable to late-stage diagnosis. Conventional visual oral examination, while widely practiced, demonstrates significant inter-examiner variability and limited sensitivity for early-stage lesions. Artificial intelligence (AI)\u0026ndash;assisted oral imaging systems represent a transformative technological advancement with the potential to standardize, accelerate, and enhance the accuracy of OSCC screening at the population level.\u003c/p\u003e\u003cp\u003e\u003cb\u003eObjectives\u003c/b\u003e\u003c/p\u003e \u003cp\u003eThis cross-sectional study aimed to evaluate the diagnostic performance of an AI-assisted oral imaging platform (OralScanNet) for early detection of OSCC, compare its performance against conventional clinical examination by dental practitioners, and identify clinicopathological and sociodemographic factors associated with screening outcomes.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMethods\u003c/b\u003e\u003c/p\u003e \u003cp\u003eA total of 1,112 adult participants were enrolled across three tertiary oral health centers between January 2023 and December 2024. Each participant underwent standardized oral imaging using a validated AI-assisted intraoral scanner integrated with a deep learning classification algorithm (ResNet-50 with attention module), followed by conventional clinical examination and histopathological biopsy as the reference standard. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC) were computed. Multivariable logistic regression identified independent predictors of OSCC detection.\u003c/p\u003e\u003cp\u003e\u003cb\u003eResults\u003c/b\u003e\u003c/p\u003e \u003cp\u003eAmong 1,112 participants, 312 (28.1%) were confirmed OSCC-positive by biopsy. The AI-assisted system demonstrated sensitivity of 92.8%, specificity of 91.3%, PPV of 84.2%, NPV of 96.1%, and AUC of 0.95 (95% CI: 0.93\u0026ndash;0.97). Conventional clinical examination achieved sensitivity of 76.4%, specificity of 82.1%, and AUC of 0.81 (95% CI: 0.78\u0026ndash;0.84). The AI system significantly outperformed conventional examination (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Tobacco use (OR 4.21; 95% CI: 2.87\u0026ndash;6.19), alcohol consumption (OR 2.94; 95% CI: 1.98\u0026ndash;4.37), and HPV-16 positivity (OR 5.73; 95% CI: 3.41\u0026ndash;9.63) were independent predictors of OSCC on multivariable analysis.\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusions\u003c/b\u003e\u003c/p\u003e \u003cp\u003eAI-assisted oral imaging demonstrated superior diagnostic accuracy compared to conventional clinical examination for early OSCC detection. Integration of AI-driven oral scanner technology into routine dental practice and community screening programs has the potential to substantially reduce late-stage diagnosis, improve survival outcomes, and enhance oral cancer screening equity across resource-varied settings. Prospective multicenter validation studies are warranted.\u003c/p\u003e","manuscriptTitle":"Artificial Intelligence–Assisted Screening for Early Detection of Oral Squamous Cell Carcinoma Using Oral Imaging Systems: A Cross-Sectional Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-13 05:57:48","doi":"10.21203/rs.3.rs-9683384/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":"fc1da6b6-5c60-46b4-9ca9-ef22c35b6290","owner":[],"postedDate":"May 13th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":67957380,"name":"Health Policy"},{"id":67957381,"name":"Dentistry"},{"id":67957382,"name":"Public Administration"},{"id":67957383,"name":"Public Relations"}],"tags":[],"updatedAt":"2026-05-13T05:57:48+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-13 05:57:48","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9683384","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9683384","identity":"rs-9683384","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}