A histopathology image–based computer-aided classification study for oral squamous cell carcinoma

Introduction: Oral squamous cell carcinoma histopathological image classification is often challenged by staining variations and sparse local lesions, which can cause models to overfit color differences and weaken cross-domain generalization. Methods: A classification framework combining staining-bias suppression and structured multiple-instance aggregation was developed. In representation learning, stain-related features were disentangled from morphological and structural information, and a gated suppression mechanism was introduced to reduce color interference while enhancing tissue architecture and cellular morphology cues. In decision aggregation, image patches were treated as instances and spatial priors were incorporated to capture both neighborhood continuity and long-range dependencies. Results: The proposed method achieved Acc 87.35%, F1 91.27%, and AUC 98.04% on one test set, and Acc 79.34%, F1 86.86%, and AUC 90.74% on another test set. It consistently outperformed traditional and deep learning baselines. External validation on an independent retrospective clinical cohort from a local hospital also showed stable performance. Discussion: The results indicate that the proposed method can effectively alleviate the impact of staining bias and improve classification robustness. Its strong performance on external data further supports its practical value under real-world acquisition and staining variations.