White Matter Hyperintensity and Multi-region Brain MRI Segmentation Using Convolutional Neural Network

Abstract

Accurate segmentation of WMH (white matter hyperintensity) from the magnetic resonance image is a prerequisite for many precise medical procedures, especially for the diagnosis of vascular dementia. Brain segmentation has important research significance and clinical application prospects especially for early detection of Alzheimer’s disease. In order to effectively perform accurate segmentation according to the MRI characteristics of different regions of the brain, this thesis proposed an optimized 3D u-net and used WHM segmentation as a pre-experiment to select the good hyperparameters (i.e. network depth, image fusion method, and the implementation of loss function) to construct an image feature learning network with both long and short skip connections. Soft voting is used as the postprocessing procedure. Our model is evaluated by a 10-fold cross-validation and achieved a dice score of 0.78 for binary segmentation (WMH segmentation) and accuracy of 0.96 for multi-class segmentation (139 regions brain segmentation), outperforming other methods.