Ensuring medical AI safety: interpretability-driven detection and mitigation of spurious model behavior and...

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computer Vision

Background:

• Deep neural networks (DNNs) are vital in medical applications but prone to shortcut learning from spurious correlations.
• Existing methods often address bias detection or mitigation separately, not comprehensively.
• Addressing these biases typically demands extensive expert labeling.

Purpose of the Study:

• To review and enhance the Reveal2Revise bias mitigation framework.
• To integrate semi-automated interpretability-based bias annotation capabilities.
• To improve the robustness and reliability of DNNs in medical tasks.

Main Methods:

• Enhanced Reveal2Revise framework with sample- and feature-level bias annotation.
• Utilized interpretability methods for semi-automated bias identification.
• Tested on four medical datasets across two modalities with controlled and real-world biases.
• Applied to VGG16, ResNet50, and Vision Transformer models.

Main Results:

• Successfully identified and mitigated biases in DNNs.
• Demonstrated improved model robustness and applicability for real-world medical tasks.
• Validated the framework's effectiveness on diverse medical data.

Conclusions:

• The enhanced Reveal2Revise framework effectively addresses shortcut learning in medical AI.
• Semi-automated bias annotation reduces reliance on extensive expert labeling.
• The approach increases the safety and trustworthiness of DNNs in clinical settings.