Interpretable AI for bio-medical applications.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Informatics

Background:

• Deep neural networks (DNNs) achieve high accuracy in medical diagnoses but often lack interpretability.
• Understanding the factors influencing DNN predictions is crucial for clinical trust and adoption.
• Explainability tools are essential for demystifying complex model behaviors.

Purpose of the Study:

• To apply Local Interpretable Model-Agnostic Explanations (LIME) and Shapley Additive exPlanations (SHAP) to interpret a DNN's breast cancer predictions.
• To compare the insights provided by LIME and SHAP in explaining individual predictions.
• To enhance the transparency of DNN models in medical applications.

Main Methods:

• A deep neural network was trained on the UCI Breast Cancer Wisconsin dataset for mass classification (benign/malignant).
• Local Interpretable Model-Agnostic Explanations (LIME) was employed to explain individual predictions.
• Shapley Additive exPlanations (SHAP) was utilized to provide both local and global explanations.

Main Results:

• Both LIME and SHAP successfully explained individual DNN predictions, highlighting key features influencing classification.
• SHAP provided a comprehensive view of feature impacts on predictions, offering a more holistic understanding.
• Commonalities in explanations derived from LIME and SHAP were identified, reinforcing the findings.

Conclusions:

• LIME and SHAP are effective tools for increasing the interpretability of deep neural networks in medical contexts.
• The application of these explainability methods enhances trust and understanding of AI-driven diagnostic tools.
• The methodology is adaptable to various DNN architectures and applications beyond breast cancer detection.