Benchmarking the influence of pre-training on explanation performance in MR image classification

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Medical Imaging
• Explainable AI (XAI)

Background:

• Convolutional Neural Networks (CNNs) are vital for medical prediction tasks, often enhanced by transfer learning for data scarcity.
• Complex CNN models lack transparency, driving the need for Explainable AI (XAI) methods.
• Quantitative evaluation of XAI performance and the impact of transfer learning remain underexplored.

Purpose of the Study:

• To propose a benchmark dataset for quantifying explanation performance in medical AI.
• To investigate the influence of transfer learning on the quality of explanations from CNN models.
• To quantitatively assess the performance of different XAI methods.

Main Methods:

• Development of a benchmark dataset for a magnetic resonance imaging (MRI) classification task.
• Application of various XAI methods to CNN models trained with and without transfer learning.
• Quantitative evaluation of explanation performance against ground-truth data.

Main Results:

• Significant performance variations exist among popular XAI methods, even for correctly classified cases.
• Explanation performance is highly dependent on the pre-training task and the number of pre-trained CNN layers.
• These findings persist after accounting for the correlation between explanation and classification performance.

Conclusions:

• Transfer learning's pre-training strategy critically influences the interpretability of medical AI models.
• The proposed benchmark enables objective assessment of XAI methods in realistic medical scenarios.
• Careful consideration of pre-training is essential for developing reliable and transparent medical AI systems.