Explaining Uncertainty in Multiple Sclerosis Lesion Segmentation Beyond Prediction Errors.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Image Analysis
• Neuroimaging

Background:

• Trustworthy artificial intelligence (AI) is crucial for healthcare, especially in medical image segmentation.
• Explainable AI and uncertainty quantification improve AI reliability, robustness, and usability.
• Limited understanding of clinical informativeness and interpretability of uncertainty in medical imaging.

Purpose of the Study:

• To introduce a novel framework for explaining predictive uncertainty sources in AI.
• To analyze uncertainty in cortical lesion segmentation for multiple sclerosis (MS) using deep ensembles.
• To shift focus from uncertainty-error to medical and engineering factors.

Main Methods:

• Developed a novel framework to explain predictive uncertainty in AI.
• Utilized deep ensembles for cortical lesion segmentation in MS.
• Analyzed instance-wise uncertainty in relation to lesion characteristics.
• Incorporated expert rater feedback.

Main Results:

• Predictive uncertainty is strongly correlated with lesion size, shape, and cortical involvement.
• Factors influencing AI uncertainty also affect human annotator confidence.
• The framework demonstrated utility across in-domain and distribution-shift scenarios.

Conclusions:

• The proposed framework effectively explains sources of predictive uncertainty in medical AI.
• Understanding uncertainty drivers like lesion size and shape enhances clinical interpretability.
• This approach advances trustworthy AI in neuroimaging and other medical applications.