Oral Cavity Anatomical Site Image Classification and Analysis.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Computational pathology

Background:

• Oral cavity cancer presents significant challenges in diagnosis and treatment, often leading to severe functional impairments and high mortality rates.
• Accurate histopathological diagnosis is crucial, necessitating precise identification of lesion locations for biopsy.
• Automated visual evaluation of oral lesions using deep learning remains an underexplored area, despite the potential for improved diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate an automated deep learning system for classifying oral cavity lesions across different anatomical sites.
• To improve the accuracy and efficiency of computer-assisted screening for oral cancer.
• To provide insights into the decision-making process of deep learning models for oral cancer detection.

Main Methods:

• Utilized the ResNeSt deep learning network, incorporating channel-wise attention and multi-path representation.
• Developed an automated method to generate labels for six distinct anatomical sites within the oral cavity.
• Analyzed model behavior using class activation maps to correlate highlighted regions with expert-identified areas of interest.

Main Results:

• Achieved an average F1-score of 0.96 across all classified anatomical sites.
• Demonstrated high accuracy (0.96) in classifying oral cavity lesion images.
• Class activation maps showed strong correlation with expert human observers' identification of regions of interest.

Conclusions:

• The ResNeSt network effectively classifies oral cavity lesion images by anatomical site, achieving high performance.
• Automated site identification is a critical first step for subsequent lesion detection modules in computer-assisted screening.
• The insights gained from model analysis can guide further improvements in AI-driven oral cancer screening tools.