High-Accuracy Oral Squamous Cell Carcinoma Auxiliary Diagnosis System Based on EfficientNet

Area of Science:

• Computational histopathology
• Artificial intelligence in oncology
• Digital pathology for cancer grading

Background:

• Manual grading of oral squamous cell carcinoma (OSCC) biopsy slides is time-consuming and subjective.
• Deep learning offers potential for automated and objective analysis of histopathological images.
• Accurate tumor grading is crucial for OSCC diagnosis and treatment planning.

Purpose of the Study:

• To develop and validate an automated deep learning system for OSCC tumor grading.
• To compare the impact of different image resolutions (10x, 20x, 40x) on diagnostic accuracy.
• To assess the system's performance on both The Cancer Genome Atlas (TCGA) and independent datasets.

Main Methods:

• Utilized EfficientNet deep learning architecture for model development.
• Trained and verified the model on a combined dataset of 1034 OSCC tissue slides (TCGA and independent).
• Evaluated diagnostic accuracy across different tile resolutions (10x, 20x, 40x).

Main Results:

• The AI system achieved 98.1% accuracy and 0.998 AUROC on the TCGA dataset.
• Optimal performance was observed at 20x resolution, with 93.1% accuracy on test tiles.
• The system demonstrated strong generalization, achieving 91.4% accuracy on an independent dataset at 20x.

Conclusions:

• The developed deep learning system provides an accurate and efficient method for OSCC grading.
• 20x resolution is optimal for automated OSCC slide analysis using this AI system.
• This computational histopathology tool can assist oral pathologists, reduce workload, and improve diagnostic efficiency.