Oral squamous cell carcinoma diagnosis: an interpretable deep learning approach using multi-scale pathological images

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Oral cancer diagnostics

Background:

• Oral squamous cell carcinoma (OSCC) diagnosis relies on subjective pathologist interpretation, leading to time-consuming processes and potential inconsistencies.
• Current diagnostic methods for OSCC face challenges with multi-center data heterogeneity and accurately identifying minority class samples.

Purpose of the Study:

• To develop a novel, accurate, and interpretable deep learning model for classifying OSCC pathological images across various magnifications.
• To address challenges in multi-center data heterogeneity and minority class sample recognition in OSCC diagnostics.

Main Methods:

• Proposed Pathology-prior Guided Multi-magnification Adaptive Fusion Network (PGMA-Net) incorporating pathological prior maps for feature attention.
• Utilized Pathology-Guided Adaptive Fusion Module (PGAFM) for multi-dimensional feature learning and Cross-Magnification Feature Alignment Regularization (CMFAR) for inter-magnification recognition.
• Implemented Dynamic Weighted Class Balance (DWCB) for minority class learning and domain adaptation for multi-center data discrepancies.

Main Results:

• PGMA-Net achieved high performance on internal test sets with an AUC of 0.9740 and accuracy of 0.9268.
• Demonstrated robust cross-center generalization on an external test set with an AUC of 0.9344 after domain adaptation.
• Model visualization confirmed alignment between its decision regions and pathologists' diagnostic bases.

Conclusions:

• PGMA-Net provides efficient and accurate classification of OSCC pathological images from diverse magnifications and multi-center data.
• The model shows strong generalization and interpretability, offering valuable assistance for pathologists in OSCC diagnosis.