Assess whether clinicians find causal counterfactual images useful for model auditing

Assess whether clinicians find counterfactual images generated from classifier-derived causal evidence useful for auditing medical image classifiers.

Background

The paper proposes a non-generative method that constructs medical-image counterfactuals from causal explanations extracted from the classifier itself. The method is evaluated quantitatively using success and similarity metrics, but the paper does not evaluate whether clinicians consider the resulting counterfactual images useful in practical model-auditing workflows. This unresolved empirical question is identified as a direction for future work.