Combining Residual learning and U-Net for Hippocampus Segmentation of Brain MRI Volume Image

Abstract

In the volume image of brain MRI, the volume of hippocampus is small, the boundary between hippocampus and surrounding tissue is fuzzy, and the two-dimensional semantic segmentation network is difficult to accurately segment. In this paper, an algorithm is proposed which combines deep residual learning and U-net for hippocampus segmentation of brain MRI volume image. It can make full use of the three-dimensional spatial information of MRI image itself, improve the ability of automatic and precise extraction of image features, and achieve high-precision hippocampus segmentation of MRI volume image. Firstly, in order to efficiently utilize 3d contextual information of the image and the solve class imbalance issue, the patches were extracted from brain MRI volume image and put into network. Then, the segmentation model based on the combination of depth residual learning and U-net is used to extract the features of image patches. After that, the upper sampling feature map and the residual learning feature map are fused to get the volume segmentation results. Finally, the detection experiments on ADNI dataset show that DSC (dice similarity coefficient) can reach 0.8915, which is better than the traditional segmentation method.