Segmenting white matter hyperintensities in brain magnetic resonance images using convolution neural networks

White matter hyperintensities (WMHs) are found on magnetic resonance (MR) images of older individuals and are associated with many neurodegenerative disorders, although the exactrole of WMHs in Alzheimer's disease and other dementias remains an open area of research. Fluid-attenuated inversion recovery (FLAIR) MR imaging sequences show WMHs with good image contrast. Manual segmentation of WMHs on FLAIR images is the widely accepted "gold standard", however, this step is often time-consuming and has a high inter-rater variability. The absence of an automated, robust and accurate approach to segment WMHs remains a processing bottleneck. We explored convolutional neural networks (CNNs) for performing semantic segmentation of WMHs in FLAIR images. Two sets of experiments were conducted: (1) Variations of U-shaped CNNs (U-Nets) were evaluated in 186 individuals, specifically, four architectures (VGG16, VGG19, ResNet152 and EfficientNetB0) having three dimensionalities (2D, 2.5D and 3D). (2) New data from 60 individuals were added to test the generalizability of U-Net, LinkNet and Feature-Pyramid Network (FPN) variants. The first experiment showed that the 2.5D implementation with VGG16 or VGG19 was the most suitable configuration when segmenting WMH (F-measure > 95% and intersection-over-union > 90%). The second experiment confirmed generalizability of these variants when using unprocessed FLAIR images.