A novel explainable AI framework for medical image classification integrating statistical, visual, and rule-based...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning models excel in healthcare data analysis but lack transparency due to their "black-box" nature.
• Existing explainable AI (XAI) methods offer limited interpretability through visualization or rule-based systems alone.
• Interpreting AI decisions is crucial for high-stakes medical applications.

Purpose of the Study:

• To develop a novel XAI method for medical image analysis that integrates statistical, visual, and rule-based explanations.
• To enhance the transparency and interpretability of deep learning models in medical image classification.
• To provide clinicians with deeper insights into AI-driven diagnostic processes.

Main Methods:

• A custom Mobilenetv2 model extracts deep features from medical images.
• A two-step feature selection (zero-based filtering and mutual importance selection) refines extracted features.
• Decision tree and RuleFit models generate human-readable rules, complemented by a novel statistical feature map overlay visualization (mean, skewness, entropy).

Main Results:

• The proposed XAI method was validated across five diverse medical imaging datasets (COVID-19, breast cancer, brain tumors, lung/colon cancer, glaucoma).
• The integrated approach provided localized and quantifiable visual explanations, enhancing model transparency.
• Results were confirmed by medical experts, indicating practical utility.

Conclusions:

• The novel XAI method significantly improves the interpretability of deep learning models in medical image classification.
• The integration of statistical, visual, and rule-based explanations offers a more comprehensive understanding of AI decisions.
• This approach holds promise for increasing trust and adoption of AI in clinical settings.