MERIT: Multi-view evidential learning for reliable and interpretable liver fibrosis staging

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Medicine

Background:

• Accurate liver fibrosis staging via MRI is clinically vital.
• Conventional methods and existing multi-view learning approaches lack uncertainty quantification and interpretability.
• Black-box feature integration in prior models compromises reliability.

Purpose of the Study:

• To introduce MERIT, a novel multi-view learning framework for liver fibrosis staging.
• To address limitations in uncertainty quantification and interpretability of existing methods.
• To enhance the reliability and clinical utility of MRI-based fibrosis staging.

Main Methods:

• Developed a multi-view method named MERIT based on evidential learning.
• Integrated uncertainty quantification using subjective logic theory for each view's prediction.
• Employed a logic-based, feature-specific combination rule for enhanced interpretability and fusion.
• Introduced a distribution-aware base rate to improve performance with class distribution shifts.

Main Results:

• MERIT demonstrated effectiveness in liver fibrosis staging.
• The method successfully quantified prediction uncertainty, enhancing reliability.
• MERIT provided both ad-hoc and post-hoc interpretability, clarifying decision-making processes.
• The significance of individual views in staging was elucidated.

Conclusions:

• MERIT offers a reliable and interpretable solution for MRI-based liver fibrosis staging.
• The framework enhances clinical decision-making through quantified uncertainty and transparent feature fusion.
• MERIT represents a significant advancement in applying multi-view learning and evidential reasoning to medical image analysis.