Interpretable deep learning approach for oral cancer classification using guided attention inference network.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Deep learning for cancer diagnostics

Background:

• Convolutional Neural Networks (CNNs) show promise for automated cancer lesion classification.
• However, CNNs often lack interpretability and can focus on irrelevant image areas, hindering reliability in biomedical applications.

Purpose of the Study:

• To develop a deep learning training approach that enhances the understandability of CNN predictions.
• To guide CNNs to accurately focus on and delineate cancerous regions in images.

Main Methods:

• Utilized gradient-weighted class activation mapping for interpretability.
• Employed a two-stage training process with data augmentation and a Guided Attention Inference Network (GAIN).
• The GAIN architecture integrates classification, attention mining, and bounding box streams for joint optimization.

Main Results:

• The developed attention maps enable understanding of the network's decision-making process.
• The method successfully guides CNNs to focus on correct oral lesion areas, improving accuracy.
• Demonstrated improved segmentation accuracy by using attention maps as reliable priors.

Conclusions:

• The proposed approach enhances the interpretability and reliability of CNNs for oral potentially malignant and malignant lesion classification.
• This method offers a more trustworthy tool for automated cancer detection in medical imaging.