A large-scale assessment of modern deep learning architectures for multiclass oral histopathology: accuracy, efficiency, and statistical reliability

Background Reliable comparison of deep learning models for oral histopathology is hindered by morphological overlap between diagnostic classes and by substantial variation in data partitioning, optimization, and evaluation across published studies. We therefore performed a large controlled benchmark to examine predictive performance, training stability, statistical uncertainty, and computational efficiency under a unified framework. Methods Thirty-one architectures from nine model families were evaluated on the three-class ORCHID and P-NDB-UFES datasets using a common downstream training, checkpoint-selection, and testing protocol. ORCHID was repartitioned at the patient/tissue-slide-group level, while the predefined 627-patch P-NDB-UFES test subset was retained unchanged because complete patient provenance was unavailable. Leading models were repeated across five random seeds. ORCHID uncertainty was estimated by cluster-aware resampling over 14 independent test groups. Evaluation included macro-F1, AUROC, AUPRC, MCC, class-specific errors, parameter count, GFLOPs, GPU forward-pass efficiency, and qualitative Grad-CAM analysis. Results MobileViTv2–150 achieved the highest primary ORCHID macro-F1 (0.9428), whereas DeiT III-Base/16 led P-NDB-UFES (0.9037). Across five seeds, the highest mean macro-F1 shifted to Swin-Large on ORCHID (0.9348 ± 0.0110) and DeiT III-Small/16 on P-NDB-UFES (0.9031 ± 0.0091). Cluster-aware ORCHID intervals overlapped among the leading models, and several compact or intermediate-scale networks matched substantially larger architectures with lower computational demand. Residual errors concentrated at morphologically difficult class boundaries and included high-confidence misclassifications. Conclusions No architecture consistently dominated both diagnostic tasks. Model selection should emphasize class-balanced performance, repeated-run stability, independence-aware uncertainty, computational cost, and clinically relevant failure patterns rather than a single-run ranking.