Automated Differentiation of Oral Red-White Lesions: An Interpretable Deep Learning Approach Combining Ensemble Architectures and Saliency Maps.

INTRODUCTION

Oral Potentially Malignant Disorders (OPMDs), including Leukoplakia and Erythroplakia, carry significant risks of malignant transformation. Early differentiation from confounding inflammatory conditions like Oral Lichen Planus (OLP) and Candidiasis is critical yet challenging due to visual similarities. This study develops a robust, interpretable deep learning framework for automated multi-class classification of pre-localized oral lesions.

MATERIALS AND

METHODS

A curated dataset of 705 high-resolution images across five categories (Leukoplakia, Erythroplakia, OLP, Candidiasis, and Normal) was utilized. We proposed a Multi-Architecture Weighted Ensemble Framework integrating ResNet-50, Xception, and EfficientNet-B0. A stratified patient-level splitting strategy prevented data leakage. Model interpretability and clinical utility were assessed via Gradient-weighted Class Activation Mapping (Grad-CAM) and Decision Curve Analysis (DCA).

RESULTS

The ensemble model achieved 91.2% accuracy and a 90.8% macro-averaged F1-score, significantly outperforming individual baselines. The strategy improved OLP detection (F1-score: 0.83), effectively distinguishing it from Leukoplakia. Grad-CAM confirmed the model focuses on pathognomonic lesion features rather than confounding artifacts. DCA suggested a potential theoretical net clinical benefit over default strategies.

CONCLUSION

This weighted ensemble framework demonstrates high retrospective accuracy and provides transparent visual explanations for the classification of pre-localized oral lesions. However, it must be interpreted strictly as a preliminary proof-of-concept investigation. While the current results suggest potential adjunctive value, extensive external validation, prospective testing, and clinician-in-the-loop studies are strictly necessary to validate its true clinical utility and impact on patient outcomes in primary care settings.