A Statistically Validated and Explainable Deep Learning Framework for Oral Cancer Classification

Despite numerous deep learning models reporting high accuracy for oral cancer classification, most rely on a single train/test split on a single dataset, with limited statistical validation, external generalization testing, or explainability. This study addresses that gap through a statistically validated, externally tested, and dual-explainable classification framework. We provide a rigorous benchmarking study contrasting 20 hybrid combinations (4 pre-trained backbones × 5 classical classifiers) against fine-tuned end-to-end CNNs. Across repeated stratified fivefold cross-validation with 10 repetitions, the hybrid framework consistently outperformed end-to-end CNNs, with ResNet50 paired with an SVM-RBF classifier achieving top performance: 0.960 accuracy (95% CI 0.958-0.963), 0.953 F1-score (95% CI 0.951-0.956), and 0.991 ROC-AUC (95% CI 0.990-0.993), with Wilcoxon signed-rank tests confirming statistical significance (p < 0.001). MobileNetV2 paired with SVM-RBF emerged as an efficient lightweight alternative, suited to resource-constrained point-of-care settings. On an independent external validation dataset, the optimal ResNet50 + SVM-RBF model demonstrated robust generalization, achieving 0.916 accuracy (95% CI 0.892-0.939), 0.984 ROC-AUC (95% CI 0.975-0.992), and 0.98 precision (95% CI 0.965-1.000), indicating a low false-positive rate on unseen samples. To ensure clinical trust, a dual Explainable AI (XAI) framework combining Grad-CAM and SHAP was implemented, providing localized visual heatmaps and global feature attribution. The proposed pipeline offers a statistically validated, generalizable, and interpretable computational framework, representing a rigorous foundation toward automated oral cancer screening support.