Here Comes the Explanation: A Shapley Perspective on Multi-contrast Medical Image Segmentation

Shapley values can reveal hidden biases in medical AI models: The study found that U-Net showed bias toward over-weighting certain MRI contrasts (T1 and FLAIR), while Swin-UNETR distributed attention more evenly across contrasts—biases that traditional pixel-level explanations would have missed, with potential implications for diagnostic accuracy.

Existing explainability methods fall short for complex medical imaging: Previous gradient-based and perturbation-based explanation techniques produce sparse, pixel-level insights that lack clinical relevance for multi-contrast MRI tasks, highlighting a gap between current XAI approaches and real-world medical needs.

Contrast-level explanations are more clinically meaningful than pixel-level ones: By explaining model decisions at the level of imaging contrasts (rather than individual pixels), this approach provides insights that radiologists and clinicians can more readily understand and act upon in clinical practice.

Different model architectures make different trade-offs in how they use data: The comparison between U-Net and Swin-UNETR demonstrates that architectural choices lead to fundamentally different decision-making patterns, suggesting that model selection should consider not just accuracy but also how transparently and fairly models use available information.

Shapley values offer a principled framework for medical AI accountability: By using game theory-based Shapley analysis, this work provides a mathematically rigorous method for auditing medical AI systems, supporting the accountability and transparency needed for clinical deployment.