Voxel-Wise Logistic Regression and Leave-One-Source-Out Cross Validation for white matter hyperintensity segmentation

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Automated segmentation of white matter hyperintensities (WMH) in brain MRI is crucial for neurological disease assessment.
• Current algorithms lack broad adoption due to inconsistent validation and performance comparison on diverse datasets.
• Existing validation frameworks may not accurately reflect real-world performance across different imaging scanners.

Purpose of the Study:

• To introduce a robust cross-validation framework, Leave-One-Source-Out Cross Validation (LOSO-CV), for evaluating WMH segmentation algorithms.
• To develop and validate a novel, accurate, and interpretable FLAIR-only WMH segmentation algorithm.
• To provide a standardized method for comparing WMH segmentation techniques on heterogeneous data.

Main Methods:

• Developed Leave-One-Source-Out Cross Validation (LOSO-CV) to assess algorithm performance on unseen scanner data.
• Introduced Voxel-Wise Logistic Regression (VLR), a FLAIR-only algorithm inspired by the Lesion Prediction Algorithm (LPA).
• Utilized a dataset of 96 brain MRI images from 7 scanners, including data from recent WMH segmentation competitions.

Main Results:

• LOSO-CV yielded more realistic (lower) performance estimates for WMH segmentation algorithms.
• The VLR algorithm achieved a median Similarity Index of 0.69, outperforming its predecessor LPA (0.58).
• VLR demonstrated improved accuracy and parameter interpretability in WMH segmentation.

Conclusions:

• LOSO-CV offers a more reliable method for validating WMH segmentation algorithms across diverse datasets and scanners.
• The VLR algorithm represents a significant advancement in FLAIR-only WMH segmentation, offering enhanced performance and interpretability.
• Standardized validation frameworks are essential for the clinical translation of automated neuroimaging analysis tools.