Automated lesion segmentation with BIANCA: Impact of population-level features, classification algorithm and locally...

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• White matter hyperintensities (WMH) show significant variability, complicating their detection.
• Factors like age, vascular issues, and neurodegeneration influence WMH characteristics.
• Existing tools like FSL's BIANCA face challenges in accurately segmenting WMH due to this variability.

Purpose of the Study:

• To enhance BIANCA, a tool for segmenting WMH, by addressing variability.
• To improve both the classification and thresholding stages of WMH segmentation.
• To develop an automated and adaptive thresholding method for more accurate WMH detection.

Main Methods:

• Incorporated population-level lesion probabilities, considering factors like age, into the classification stage.
• Evaluated alternative machine learning classifiers against the existing K-nearest neighbor algorithm.
• Introduced LOCally Adaptive Threshold Estimation (LOCATE) as a supervised method for adaptive thresholding, contrasting it with global thresholding.

Main Results:

• Including population-level probabilities and alternative classifiers yielded minimal improvements.
• LOCATE demonstrated substantial gains in WMH segmentation performance over global thresholding.
• LOCATE improved detection of deep lesions and segmentation of periventricular boundaries across diverse datasets.

Conclusions:

• LOCATE significantly enhances WMH segmentation accuracy by adapting to local lesion probabilities.
• The method shows robustness across different datasets and patient cohorts, including those with CADASIL.
• LOCATE offers a superior alternative to global thresholding for automated WMH segmentation.