Counterfactual Reasoning for Mammogram Classification via Semantic Texture Masking

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Artificial intelligence-based computer-aided diagnosis (CADx) systems are increasingly used in mammography.
• Limited interpretability of AI decision-making hinders clinical trust.
• Understanding model reliance on lesion versus surrounding tissue is crucial for improving CADx systems.

Purpose of the Study:

• To investigate whether deep learning classifiers focus on lesion characteristics or surrounding breast tissue.
• To assess model interpretability and robustness using counterfactual reasoning with semantic masking.

Main Methods:

• Modified mammogram textures by selectively removing information from lesion (foreground) or non-lesion (background) regions.
• Trained and evaluated MobileNet, ResNet50, and ResNet50v2 on the CBIS-DDSM dataset.
• Assessed classification performance using the area under the ROC curve (AUC) across four masking scenarios.

Main Results:

• All models performed similarly on unaltered mammograms.
• ResNet50 showed significant performance degradation when background information was removed, indicating high dependence on context.
• ResNet50v2 demonstrated improved robustness compared to ResNet50, while MobileNet was stable across all masking scenarios.

Conclusions:

• Deep learning models exhibit different dependencies on lesion versus background information in mammograms.
• ResNet50v2 and MobileNet show greater robustness, suggesting better preservation of lesion-specific features.
• Understanding region-specific dependencies enhances AI interpretability and aids in developing reliable clinical CADx systems.