Abstract
Oral submucous fibrosis (OSMF) is a chronic, progressive, potentially malignant disorder prevalent in Southeast Asia and the Indian subcontinent. Progressive fibrosis restricts mouth opening, making interincisal opening as a key parameter of disease staging. This study developed a portable AI-based device for automated interincisal opening measurement and OSMF lesion detection important for real-time screening and staging. A portable device powered by a quad-core ARM Cortex-A72 processor integrates a YDLiDAR depth camera for interincisal opening measurement and an intraoral camera for image acquisition. Structured inquiry into temporomandibular joint symptoms and OSMF-associated habit history reduces false-positive classification. The AI pipeline comprises buccal mucosa segmentation, OSMF classification, Stages I-III grading based on interincisal opening, and lesion detection. Model robustness is evaluated using stratified fivefold cross-validation and Grad-CAM analysis. Bland-Altman analysis assesses agreement between LiDAR and vernier-caliper measurements, while quadratic weighted Cohen's kappa evaluates agreement between AI- and clinician-assigned OSMF stages. A total of 350 OSMF (including 10 biopsy-confirmed cases) and 334 non-OSMF images were collected. U-Net with EfficientNet-B2 achieved accurate mucosal segmentation, DenseNet-121 showed the best classification performance (accuracy/F1, 0.991), and U-Net with ResNet-34 enabled lesion localization. Clinical validation in 100 subjects achieved 98.0% sensitivity, 96.0% specificity, and 97.0% accuracy. Bland-Altman analysis confirmed agreement between interincisal and vernier-caliper measurements (SD, 0.68 mm), while AI-based staging showed substantial agreement (weighted kappa k = 0.982, 95% CI 0.956-1.00). The proposed system provides a reliable, cost-effective solution for early OSMF screening and decision support in government healthcare and community settings.