3D Deep Learning Segmentation of White Matter Hyperintensity on 3T and 7T Brain MRI Scans

White matter lesions (WMLs), commonly found as hyperintensities (WMHs) on T2-weighted FLAIR MR brain images, are associated with neuropsychiatric and neurodegenerative disorders. In the present study, we adapted a 3D U-Net deep learning method to automatedly segment the WMHs on 3T and 7T MRI T2w FLAIR brain images. Using 3D U-Net, the accuracy of WMH segmentation is 98.8% for 3T data, while it drops to 90.3% for 7T data. However, after incorporating histogram matching in the preprocessing, the accuracy of WMHs segmentation significantly improves to 97.5% for 7T data.