Localization Gains from Explanation Regularization Should Not Be Interpreted as Faithfulness Gains.

Overview & Synopsis

Training-time explanation regularization is often evaluated by whether saliency maps become more concentrated within annotated regions. We argue that this criterion alone is insufficient: transformation consistency, lesion localization, and faithfulness to the classifier are distinct properties and should be evaluated separately, particularly in medical imaging. We examine this distinction through a case study of Contrastive Grad-CAM Consistency (CGC) in three-class brain tumor MRI classification. Using a single-channel ResNet-50 with five-fold patient-grouped cross-validation, we compare a cross-entropy baseline with CGC. CGC changes overall Pointing Game (PG) from $27.24\%$ to $36.02\%$ and Relevance Mass Accuracy (RMA) from $11.17\%$ to $18.01\%$, while classification accuracy changes from $93.63\%$ to $94.87\%$. Perturbation-based faithfulness metrics show inconsistent changes: Insertion Area Under the Curve (IAUC) changes from $0.73$ to $0.78$, whereas Deletion Area Under the Curve (DAUC) favors the baseline ($0.36$ versus $0.41$; lower is better), and overall Iterative Removal Of Features (IROF) is $0.57$ for both models. These results support our position that improved spatial agreement with pathology should not be interpreted as a general improvement in explanation faithfulness. We therefore recommend property-specific evaluation in which the property targeted by the explanation objective, spatial agreement with reference regions, and model reliance are measured and reported separately.