Detecting and Mitigating the Clever Hans Effect in Medical Imaging: A Scoping Review

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• The Clever Hans effect, or shortcut learning, occurs when AI models exploit spurious correlations instead of genuine clinical features.
• This phenomenon poses a significant barrier to developing trustworthy artificial intelligence (AI) in medical imaging.
• Reliable AI systems are crucial for advancing diagnostic and prognostic capabilities in healthcare.

Purpose of the Study:

• To conduct a scoping review of methods for identifying and addressing the Clever Hans effect in medical imaging AI.
• To classify existing detection and mitigation strategies for shortcut learning.
• To highlight the need for improved validation and transparency in AI research.

Main Methods:

• A comprehensive literature search was performed for papers published between 2010 and 2024.
• 173 papers were reviewed, with 37 selected for detailed analysis.
• Methods were categorized into detection (model-centric, data-centric, uncertainty/bias-based) and mitigation (data manipulation, feature disentanglement, domain knowledge).

Main Results:

• Detection methods focus on identifying reliance on non-clinical patterns.
• Mitigation strategies aim to remove or correct these spurious correlations.
• A significant gap exists as most studies do not report or test for shortcut learning, indicating a need for more rigorous validation.

Conclusions:

• Despite advancements, the Clever Hans effect remains a challenge in medical imaging AI.
• Increased transparency, standardized benchmarks, and automated detection tools are necessary.
• Community-driven best practices and interdisciplinary collaboration are vital for developing reliable and equitable AI in healthcare.