Improving Explainability of Disentangled Representations using Multipath-Attribution Mappings

Disentangled representations go beyond pixel-level explanations: Traditional explainability methods only highlight which pixels matter for a diagnosis, but this paper shows that breaking down a model's internal representations into interpretable components allows experts to understand why certain visual features are being used—a crucial distinction for clinical safety.

The framework enables causal reasoning in medical AI: By visualizing these disentangled representations, domain experts can leverage their medical knowledge to investigate potential causal relationships rather than just correlations, addressing a key gap in current explainable AI approaches for high-stakes healthcare decisions.

Multi-path attribution mapping validates explanations: The paper introduces a validation technique that enriches explanations through multiple pathways, helping clinicians and regulators gain confidence that the model's reasoning is sound and not based on spurious patterns.

Shortcut detection improves model robustness without extra testing: The framework can identify when models rely on problematic "shortcuts" (like using irrelevant features) without requiring expensive testing under different data distributions—a practical advantage for deploying safer clinical AI systems.

Effectiveness demonstrated on real medical data: The approach was validated on synthetic benchmarks and two actual medical datasets, suggesting the framework has practical applicability beyond theoretical proof-of-concept for clinical diagnostic systems.