Explaining the black-box smoothly-A counterfactual approach.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Explainable AI

Background:

• Classical feature importance methods like saliency maps fail to explain *how* imaging features influence medical AI classification decisions.
• Transparent decision-making is critical for adopting AI in healthcare applications.
• Existing explanation methods lack clinical relevance and user-centric evaluation.

Purpose of the Study:

• To introduce a novel BlackBox Counterfactual Explainer for medical image classification models.
• To generate clinically interpretable explanations by progressively altering images to change classification outcomes.
• To audit a chest X-ray classifier and evaluate the clinical relevance of explanations.

Main Methods:

• Utilized a Generative Adversarial Network (GAN) to create counterfactual image perturbations.
• Developed a loss function to preserve essential anatomical details during image generation.
• Proposed clinically-relevant quantitative metrics (e.g., cardiothoracic ratio) for explanation evaluation.
• Conducted a user study with radiology residents comparing different explanation methods.

Main Results:

• The proposed counterfactual explanation method significantly enhanced user understanding of AI decisions compared to no explanation or saliency maps.
• Clinically-relevant metrics demonstrated quantifiable differences in explanations between positive and negative diagnostic predictions.
• User study revealed superior performance of counterfactual explanations in understandability, justification, and overall helpfulness.

Conclusions:

• The BlackBox Counterfactual Explainer offers a transparent and clinically relevant approach to understanding medical AI.
• The developed metrics provide a benchmark for evaluating AI explanation methods in medical imaging.
• The findings suggest AI classifiers can be audited to ensure reliance on clinically meaningful features, enhancing trust in medical AI.