Explainable AI framework for detecting and grading oral squamous cell carcinoma in South Asian populations using histopathological images

Abstract

Oral squamous cell carcinoma (OSCC) represents a significant public health burden in South Asia, where a critical shortage of trained oral pathologists contributes to diagnostic delays and inconsistent histopathological grading. The study aimed to develop and evaluate an interpretable deep learning based framework for automated oral cancer detection and grading from histopathological images. A two-stage transfer learning framework was developed using the publicly available ORCHID dataset, comprising histopathological images from 150 patients across five Indian institutions. Stage 1 classified images into Normal, Oral Submucous Fibrosis (OSMF), and OSCC, while Stage 2 graded confirmed OSCC cases into well-differentiated, moderately-differentiated, and poorly-differentiated subtypes. Transfer learning approach was applied on three pretrained architectures (EfficientNetB0, InceptionV3 and ResNet50) and Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to assess model interpretability. Among the three architectures, EfficientNetB0 performed best for Stage 1, achieving an accuracy of 90.37% (95% CI: 89.54–91.20), macro F1-score of 79.39% (77.25–81.53) and AUC of 0.91 (0.89–0.93), outperforming InceptionV3 (87.48% accuracy) and ResNet50 (84.88% accuracy). Stage 2 (EfficientNetB0) achieved a macro accuracy of 89.99% (88.97–91.02), macro F1-score of 84.67% (82.98–86.35) and AUC of 0.93 (0.92–0.94). Grad-CAM heatmaps showed consistent attention on histopathologically relevant regions, supporting the clinical interpretability of the model. The proposed framework demonstrated clinically meaningful performance for automated oral cancer detection and grading, with explainability analysis confirming alignment with established histopathological criteria. The findings suggest that such a framework have the potential to serve as a viable decision-support tool for ENT practice in resource-limited South Asian settings. However, further validation through multicenter prospective studies is required before clinical use.