Performance comparison of 10 different classification techniques in segmenting white matter hyperintensities in aging

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• White matter hyperintensities (WMHs) are key indicators of small vessel disease in aging and Alzheimer's disease (AD).
• Manual segmentation of WMHs is time-consuming and suffers from inter- and intra-rater variability.
• Automated tools are needed for robust WMH detection and measurement of vascular burden.

Purpose of the Study:

• To compare the performance of 10 different linear and nonlinear classification techniques for WMH segmentation from MRI data.
• To identify the optimal classifier for accurate and robust WMH detection.
• To assess the impact of different MRI contrast combinations on classifier performance.

Main Methods:

• Evaluated 10 classifiers (e.g., Random Forests, SVM, k-NN) using intensity and spatial location features from multi-site MRI datasets (ADC, NACC, ADNI).
• Assessed classifier performance using similarity indices (Dice Kappa, ICC) and computational burden.
• Compared performance with varying MRI contrasts, including T1w, T2w, PD, and FLAIR sequences.

Main Results:

• Random Forests achieved the best performance across all datasets, with high Dice Kappa and ICC scores.
• FLAIR information was crucial for optimal performance, though Random Forests maintained strong results even without T2w/PD data.
• Performance of linear classifiers and some nonlinear classifiers declined significantly without FLAIR information.

Conclusions:

• Random Forests are the most suitable classifiers for robust WMH segmentation from MRI data.
• Automated WMH detection tools, particularly those employing Random Forests, can aid in assessing vascular burden in aging and AD.
• The study provides an open-source automated tool for WMH segmentation with pre-trained classifiers.