Deep Neural Networks with Multi-scale Highlighted Foregrounds for White Matter Hyperintensities Segmentation

Abstract

White matter hyperintensities (WMHs) are abnormal signals within the white matter region on the human brain MRI and have been associated with aging processes, cognitive decline, and dementia. In the current study, we proposed a U-Net with multi-scale highlighted foregrounds (HF) for WMHs segmentation. Our method, U-Net with HF, is designed to improve the detection of the WMH voxels with partial volume effects. We evaluated the segmentation performance of the proposed approach using the WMH Segmentation Challenge dataset. We demonstrated that the U-Net with HF significantly improved the detection of the WMH voxels at the boundary of the WMHs or in small WMH clusters quantitatively and qualitatively. Up to date, the proposed method has achieved the best overall evaluation scores, the highest dice similarity index, and the best F1-score among 47 methods submitted on the WMH Segmentation Challenge that was initially hosted by MICCAI 2017 and is continuously accepting new challengers. The implementation of our proposed method is publicly available using Dockerhub (https://hub.docker.com/r/wmhchallenge/pgs). Then we assessed the clinical utility of the WMHs volumes that were automatically computed using our method using the Alzheimer’s Disease Neuroimaging Initiative database. The evaluation of the clinical utility showed that the WMHs volume that was automatically computed using U-Net with HF was significantly associated with age, group diagnosis information (cognitively normal, mild cognitive impairment, and Alzheimer's disease), and cognitive performance scores and the WMHs volume improves the classification between cognitive normal and Alzheimer's disease subjects and between patients with mild cognitive impairment and those with Alzheimer's disease. The classifier based on deep neural networks trained with the spatial distribution of WMHs extracted by our method showed notable classification performance for AD versus CN. The class activation map of AD extracted from the classifier using CAM and Grad-CAM detected a reasonable region. As a result, we confirmed that the WMHs extracted by ours were clinically useful.