Artificial Intelligence-Based Classification of Oral Pigmented Lesions in Adolescents

Abstract

Background:

Oral pigmented lesions in adolescents encompass a range of conditions with overlapping clinical features, making accurate recognition important for appropriate diagnosis and management. Artificial Intelligence (AI), particularly deep learning, may provide useful support for image-based classification of these lesions. This study evaluated the diagnostic performance of a ResNet50-based AI model in classifying oral pigmented lesions among adolescents.

Materials and Methods:

An observational diagnostic accuracy study was conducted at a tertiary care centre from January to March 2026. A total of 150 adolescents aged 12–19 years with clinically identifiable oral pigmented lesions were evaluated. Standardized intraoral photographs and relevant clinical records were assessed using a ResNet50 convolutional neural network (CNN). Lesions were classified into five diagnostic categories: melanotic macule, melanocytic nevus, physiological pigmentation, exogenous pigmentation, and melanoacanthoma. AI-generated classifications were compared with the reference diagnosis. Diagnostic performance was assessed using accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1-score, and Cohen’s kappa (κ). AI performance was also compared with that of a clinical examiner.

Results:

The AI model correctly classified 139 of 150 cases, yielding an overall accuracy of 92.7%. Category-specific accuracy ranged from 90.0% to 96.7%. The overall sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were 92.7%, 98.2%, 92.7%, and 98.2%, respectively. κ demonstrated excellent agreement between AI-based classification and the reference diagnosis (κ = 0.909; 95% confidence interval (CI): 0.846–0.972; probability (p) < 0.001). Classification accuracy remained comparable across gender and adolescent age groups. The AI model achieved higher observed accuracy than the clinical examiner, with accuracies of 92.7% and 87.3%, respectively.

Conclusion:

The ResNet50-based AI system demonstrated reliable classification performance and excellent concordance with the reference diagnosis for oral pigmented lesions in adolescents. Its consistent performance across demographic groups suggests potential utility as an adjunctive diagnostic tool. Larger multicentre studies with independent external validation and prospective clinical evaluation are warranted before broader clinical implementation.