Attention-guided convolutional network for bias-mitigated and interpretable oral lesion classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Accurate diagnosis of oral lesions is crucial for early oral cancer detection, presenting a significant clinical challenge.
• Deep learning (DL) models show promise in assisting clinical decision-making for oral lesion classification.

Purpose of the Study:

• To develop and evaluate a DL model for classifying oral lesions with improved accuracy, interpretability, and reduced dataset bias.
• To introduce a novel GAIN (Guidance and Attention) model incorporating anatomical site prediction for enhanced diagnostic support.

Main Methods:

• A CNN-based Classification Stream was used as a baseline for 16-class oral lesion categorization.
• A Guidance Stream (GAIN model) aligned class activation maps with ground truth segmentation masks for better localization.
• An Anatomical Site Prediction Stream was added (GAIN+ASP model) to further improve interpretability.

Main Results:

• The GAIN model achieved a 7.2% relative accuracy improvement over the baseline for 16-class classification.
• Superior class-specific balanced accuracy and AUC scores were observed with the GAIN model.
• The models demonstrated enhanced lesion localization, improved attention map alignment, and greater robustness against dataset bias.

Conclusions:

• The developed deep learning models, particularly GAIN and GAIN+ASP, offer a promising approach for accurate and interpretable oral lesion classification.
• These models have the potential to support clinicians in the early diagnosis of oral cancer by improving diagnostic accuracy and reducing bias.