AI-Assisted Diagnosis of Pediatric Oral Ulcerative Lesions Using Clinical and Image-Based Features

Abstract

Background:
Oral ulcerative lesions are frequently encountered in pediatric patients and may arise from traumatic, infectious, recurrent aphthous, or inflammatory conditions. Because several of these disorders exhibit overlapping clinical characteristics, establishing an accurate diagnosis can be challenging. Artificial intelligence (AI), particularly deep-learning techniques applied to clinical images, may provide valuable support for lesion recognition and diagnostic classification.
Aim:
To assess the diagnostic effectiveness of an AI-assisted approach integrating clinical and image-derived characteristics for classifying pediatric oral ulcerative lesions and to compare its performance with conventional clinical diagnosis.
Materials and Methods:
This study included 120 children aged 6–14 years presenting with clinically identifiable oral ulcerative lesions. Clinical parameters, including pain, perilesional erythema, lesion number, regional lymphadenopathy, systemic symptoms, lesion duration, and lesion size, were documented together with standardized clinical photographs. Image-derived characteristics included ulcer margin, base appearance, surrounding mucosal changes, and lesion number. Three deep-learning architectures—ResNet50, VGG16, and InceptionV3—were assessed, followed by evaluation of a proposed combined AI model. Diagnostic performance was determined using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, F1-score, receiver operating characteristic (ROC) analysis, and area under the curve (AUC). Agreement with the reference diagnosis was assessed using Cohen’s kappa coefficient, while diagnostic performance was compared with conventional clinical assessment.
Results:
Traumatic ulcers constituted the most frequent diagnostic category (25.0%), followed by recurrent aphthous ulcers (23.3%) and herpetic ulcers (18.3%). Pain and perilesional erythema were observed in 77.5% and 71.7% of participants, respectively. The proposed combined model produced the best overall performance, achieving 95.0% accuracy, 94.2% sensitivity, 96.1% specificity, 93.5% PPV, 96.5% NPV, and a 93.8% F1-score. The overall AUC was 0.978 (95% CI: 0.956–0.992). Diagnostic accuracy across individual lesion categories ranged from 94.2% to 96.7%. Agreement between the AI-generated diagnosis and the reference diagnosis was almost perfect (κ = 0.89; 95% CI: 0.83–0.95), whereas conventional clinical diagnosis demonstrated substantial agreement (κ = 0.70; 95% CI: 0.60–0.80). The AI model also achieved greater overall diagnostic accuracy than conventional clinical assessment (95.0% vs. 84.2%).
Conclusion:
The proposed AI-assisted approach demonstrated strong diagnostic capability for classifying pediatric oral ulcerative lesions and showed greater concordance with the reference diagnosis than conventional clinical assessment. Combining clinical information with image-based characteristics may enhance diagnostic support in pediatric oral healthcare. Nevertheless, validation in larger, multicenter, and more heterogeneous populations is necessary to establish the model’s generalizability and clinical applicability before routine implementation.