CONVOLUTIONAL NEURAL NETWORK-BASED BINARY CLASSIFICATION FOR ORAL EPITHELIAL DYSPLASIA

Abstract

Objective: This study aims to advance digital pathology in oral diagnostics by developing an artificial intelligence (AI) tool capable of detecting oral epithelial dysplasia (OED) and identifying its key histopathological features.

Methods: We employed a retrospective analytical design using archived hematoxylin and eosin–stained histopathological slides from previously diagnosed cases obtained from a public hospital. Slides were digitized and used to train and evaluate a convolutional neural network (CNN) for binary classification (OED vs. non-OED). Model performance was assessed using accuracy, sensitivity, specificity, precision, F1-score, and Cohen’s kappa (κ) to measure agreement with expert histopathological diagnoses.

Results: CNN achieved an overall diagnostic accuracy of 88.1%, with a sensitivity of 87.5% and specificity of 88.7%. Cohen’s κ of 0.76 indicated substantial inter-rater agreement between the AI predictions and the reference diagnoses. We have used CNN-based model, which demonstrates the strong potential for accurate, automated detection of OED in histopathological images. By enhancing diagnostic consistency and mitigating interobserver variability, this AI-driven approach may contribute to earlier detection, improved risk stratification, and better clinical outcomes in oral pathology.