Benchmarking the Influence of Pre-training on Explanation Performance in MR Image Classification

Explanation methods vary dramatically in quality: Popular explainability techniques applied to the same medical AI model produced vastly different results, meaning the choice of explanation method significantly impacts what insights doctors receive about how the model makes decisions—even when the model's predictions are correct.

Pre-training choices matter for trustworthiness: The specific task used to pre-train an AI model and how many layers were pre-trained substantially influenced the quality of explanations the model could provide, suggesting that transfer learning decisions affect not just accuracy but also interpretability and trustworthiness.

We need better evaluation standards for explanations: The researchers found that previous studies rarely quantitatively tested whether explanations actually matched ground-truth data, highlighting a gap in how the field validates that XAI methods are truly reliable before deployment in clinical settings.

Explanation quality isn't guaranteed by prediction accuracy: Even when models correctly classified MRI images, their explanations varied widely in quality, indicating that a model's correct predictions don't automatically mean its reasoning is sound or understandable to medical professionals.

Benchmarking is essential for responsible AI in healthcare: This work demonstrates the need for standardized testing methods to evaluate explanation performance in medical AI, which is critical for ensuring clinicians can trust and verify how these systems reach diagnostic conclusions.