Uncertainty quantification for White Matter Hyperintensity segmentation detects silent failures and improves...

Area of Science:

• Neuroradiology
• Medical Image Analysis
• Machine Learning

Background:

• White Matter Hyperintensities (WMH) are crucial markers of small vessel disease in brain MRI, vital for clinical and research assessments.
• Accurate WMH segmentation is challenging due to variability in size, shape, borders, and overlap with other pathologies or artifacts.

Purpose of the Study:

• To evaluate uncertainty quantification (UQ) techniques for improving WMH segmentation accuracy and reliability in brain MRI.
• To assess the utility of UQ in identifying segmentation failures and enhancing downstream clinical applications, such as Fazekas score classification.

Main Methods:

• Assessed various UQ techniques for WMH segmentation across different data distributions.
• Utilized Stochastic Segmentation Networks and Deep Ensembles for segmentation.
• Developed a novel method for Fazekas score classification using WMH probability and UQ maps.

Main Results:

• UQ techniques effectively identified previously unsegmented WMH, reducing 'silent failures'.
• Stochastic Segmentation Networks with Deep Ensembles achieved superior Dice and Absolute Volume Difference % (AVD) scores.
• Integrating UQ information improved Fazekas score classification balanced accuracy and calibration, outperforming models without UQ.

Conclusions:

• UQ is valuable for improving the robustness and reliability of WMH segmentation in brain MRI.
• UQ-enhanced spatial features significantly boost the performance of clinical scoring systems like the Fazekas score.
• Stochastic UQ methods with diverse samples can detect and mitigate poor quality segmentations.