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

Area of Science:

• Histopathology
• Medical Image Analysis
• Computational Pathology

Background:

• Oral squamous cell carcinoma (OSCC) classification faces challenges due to staining variations and sparse lesions.
• These issues can lead to model overfitting and reduced generalization across different datasets.

Purpose of the Study:

• To develop a robust classification framework for OSCC histopathological images.
• To address staining bias and improve cross-domain generalization.

Main Methods:

• A framework combining staining-bias suppression and structured multiple-instance aggregation was developed.
• Stain-related features were disentangled, and a gated suppression mechanism was used to reduce color interference.
• Spatial priors were incorporated for patch-based instance aggregation, considering neighborhood continuity and long-range dependencies.

Main Results:

• The method achieved high accuracy (87.35%) and F1-score (91.27%) on one test set, and (79.34%, 86.86%) on another.
• It consistently outperformed traditional and deep learning baselines.
• Stable performance was observed on an independent clinical cohort, demonstrating practical value.

Conclusions:

• The proposed method effectively mitigates staining bias in OSCC histopathological image classification.
• The approach enhances classification robustness and shows practical utility under real-world conditions.