Trustworthy Multi-Modal Ultrasound Fusion via Uncertainty Calibration and Conflict Resolution.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Diagnostic Ultrasound

Background:

• Multi-modal ultrasound enhances lesion assessment by integrating diverse imaging data.
• Conventional methods often assume uniform data quality, overlooking noise and patient variability, which compromises diagnostic reliability.
• Challenges include quantifying modality-specific uncertainty and resolving prediction conflicts in fusion.

Purpose of the Study:

• To develop a novel method for trustworthy multi-modal ultrasound fusion that addresses modality-wise uncertainty and prediction conflicts.
• To improve the diagnostic accuracy and robustness of multi-modal ultrasound analysis.

Main Methods:

• Proposed Trustworthy Multi-modal Ultrasound Fusion (TMUF) method.
• Introduced a cross-modal uncertainty calibration regularizer to estimate evidence-based uncertainty.
• Developed a credibility-aware fusion strategy to dynamically weight modalities based on consistency and uncertainty.

Main Results:

• TMUF achieved diagnostic accuracies of 88.00% for breast lesions and 92.08% for liver cancer.
• The method outperformed state-of-the-art baselines in validation studies.
• Demonstrated enhanced diagnostic accuracy and robustness in multi-modal ultrasound fusion.

Conclusions:

• TMUF effectively integrates information from multiple ultrasound modalities by managing uncertainty and conflicts.
• The proposed approach significantly improves diagnostic performance in challenging clinical scenarios.
• TMUF offers a robust solution for enhancing the reliability of multi-modal ultrasound diagnostics.