<p>To develop and evaluate MultiDentNet, a unified deep learning framework for multi-class dental condition screening and preliminary risk stratification of cancer-suspicious oral lesions, utilizing backbone-diverse ensembling and inter-class relational modeling. Four complementary CNN backbones (DenseNet-121, EfficientNetV2-S, ResNet-50, Inception-V3) integrated with Squeeze-and-Excitation (SE) attention, graph convolutional networks (GCNs), and multi-task learning were fused via validation-optimized weighting. The framework was evaluated on 10,235 clinical intraoral images (five conditions) and 940 clinically labeled oral lesion images (cancer-suspicious vs. non-cancer-suspicious; histopathological confirmation unavailable). Performance was assessed using accuracy, Cohen’s <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\kappa\)</EquationSource> </InlineEquation>, Matthews correlation coefficient (MCC), false negative rate (FNR), and bootstrapped 95% confidence intervals (CIs), alongside simulated domain-shift robustness testing. The ensemble achieved 99.70% accuracy (95% CI 99.32–100.00%) for dental classification (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\kappa = 0.996\)</EquationSource> </InlineEquation>) and 95.71% (95% CI 92.20–98.58%) for oral lesion risk stratification (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\kappa = 0.913\)</EquationSource> </InlineEquation>). For cancer-suspicious lesions, recall reached 97.31% (FNR: 2.69%, 95% CI 0.00–7.15%). Architectural diversity successfully mitigated class imbalance, significantly reducing the Hypodontia FNR to 1.65% (95% CI 0.00–4.24%) compared to single-model baselines. Performance demonstrated moderate resilience to acquisition variability (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\Delta\)</EquationSource> </InlineEquation>accuracy <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\approx\)</EquationSource> </InlineEquation> 8–12% degradation under brightness and contrast perturbations) but degraded substantially under high-frequency noise. Grad-CAM visualizations localized attention to clinically relevant morphological features. MultiDentNet provides an interpretable, efficient baseline for dental screening and lesion triage. Serving as an adjunctive proof-of-concept, its metrics reflect an upper bound due to reliance on single-center, clinically labeled data. Prospective multi-center validation with histopathological standards remains necessary. Code: <a href="https://github.com/Aliyar4061/MultiDentNetV3">https://github.com/Aliyar4061/MultiDentNetV3</a>. Positioned as an adjunctive triage tool for resource-constrained settings, high-volume workflows, and tele-dentistry, designed to augment clinician-in-the-loop oversight.</p>

MultiDentNet: a unified deep learning framework for multi-class dental condition screening and preliminary oral lesion triage

期刊论文
出版模式：

开放获取
发表日期：

2026年5月13日

查看全文

Ali Zeydi Abdian,
Mohammad Masoud Javidi,
Najme Mansouri,
Farzaneh Mehranfar,
Sahar Cheperli

摘要

To develop and evaluate MultiDentNet, a unified deep learning framework for multi-class dental condition screening and preliminary risk stratification of cancer-suspicious oral lesions, utilizing backbone-diverse ensembling and inter-class relational modeling. Four complementary CNN backbones (DenseNet-121, EfficientNetV2-S, ResNet-50, Inception-V3) integrated with Squeeze-and-Excitation (SE) attention, graph convolutional networks (GCNs), and multi-task learning were fused via validation-optimized weighting. The framework was evaluated on 10,235 clinical intraoral images (five conditions) and 940 clinically labeled oral lesion images (cancer-suspicious vs. non-cancer-suspicious; histopathological confirmation unavailable). Performance was assessed using accuracy, Cohen’s \(\kappa\) , Matthews correlation coefficient (MCC), false negative rate (FNR), and bootstrapped 95% confidence intervals (CIs), alongside simulated domain-shift robustness testing. The ensemble achieved 99.70% accuracy (95% CI 99.32–100.00%) for dental classification ( \(\kappa = 0.996\) ) and 95.71% (95% CI 92.20–98.58%) for oral lesion risk stratification ( \(\kappa = 0.913\) ). For cancer-suspicious lesions, recall reached 97.31% (FNR: 2.69%, 95% CI 0.00–7.15%). Architectural diversity successfully mitigated class imbalance, significantly reducing the Hypodontia FNR to 1.65% (95% CI 0.00–4.24%) compared to single-model baselines. Performance demonstrated moderate resilience to acquisition variability ( \(\Delta\) accuracy \(\approx\) 8–12% degradation under brightness and contrast perturbations) but degraded substantially under high-frequency noise. Grad-CAM visualizations localized attention to clinically relevant morphological features. MultiDentNet provides an interpretable, efficient baseline for dental screening and lesion triage. Serving as an adjunctive proof-of-concept, its metrics reflect an upper bound due to reliance on single-center, clinically labeled data. Prospective multi-center validation with histopathological standards remains necessary. Code: https://github.com/Aliyar4061/MultiDentNetV3. Positioned as an adjunctive triage tool for resource-constrained settings, high-volume workflows, and tele-dentistry, designed to augment clinician-in-the-loop oversight.