Architecture-aware Robustness Evaluation of Explainable Deep Learning for Breast Cancer Diagnosis

There is no one-size-fits-all explanation method for AI medical imaging: The study tested nine different explainability techniques across four types of AI models for breast cancer diagnosis and found that "no single technique generalises across architectures," meaning hospitals cannot simply pick their favorite explanation tool—they must carefully match it to their specific AI system.

The AI model architecture matters as much as the explanation method: The research found that explanation quality depends on "the interaction between model design and explanation strategy" rather than one factor dominating the other, suggesting that responsible AI deployment requires coordinated choices about both the underlying model and how it explains its decisions.

Hybrid AI architectures provide clearer, more trustworthy explanations: The study found that hybrid neural networks "produce more spatially coherent explanations" compared to other model types, which is important for clinical settings where doctors need to understand exactly where the AI identified potential cancer.

Some AI models are less predictable in their explanations: Lightweight and transformer-based models "exhibit greater variability across methods," meaning their explanations can be inconsistent depending on which explanation technique is used—a potential concern for clinical reliability and trust.

Systematic evaluation is necessary before deploying medical AI: Rather than ad hoc selection of explanation tools, this research demonstrates the need for rigorous testing using clinically relevant metrics to ensure that AI systems used in breast cancer diagnosis can genuinely explain their recommendations to doctors.