Objectives: To develop and evaluate a dual-stage deep learning framework for automated detection and grading of oral squamous cell carcinoma (OSCC) using whole-slide histopathology images. Methods: A multi-institutional dataset comprising 1,586 OSCC slides and 330 normal slides was used. The ORALPATHO pipeline utilized attention-based multiple instance learning for tumor detection and gated attention with top-K patch selection for subtype classification. Model performance was assessed using cross-validation and external datasets. Results: The binary classifier achieved a mean F1-score above 0.93. Multiclass grading achieved macro-F1 scores between 0.68 and 0.72. Poorly differentiated tumors showed lower classification accuracy due to limited sample representation and morphological heterogeneity. Conclusions: Attention-based weakly supervised learning is feasible for OSCC grading. The ORALPATHO framework provides clinically interpretable outputs and supports future development of AI-assisted oral cancer diagnostics.
Chaurasia et al. (Sun,) studied this question.