PRU-net: An U-net Model with Pyramid Pooling and Residual Block for WMH Segmentation

Abstract

Segmentation of white matter hyperintensities (WMHs) from MR images is an essential step in computer-aided diagnosis of brain diseases, especially when considering their effect on cognition or stroke. At present, most of the research for WMH segmentation is based on deep learning methods. Although many deep learning segmentation methods have been proposed, their accuracy of these methods still needs to be improved, especially for discrete and small-sized deep WMHs. To cope with these challenges, and to improve the accuracy of WMH segmentation, an improved 3D U-net model, named PRU-net, was proposed in this paper. PRU-net integrates pyramid pooling and residual convolutional block in bottleneck layer of the U-net architecture. The pyramid pooling block was used to aggregate more context information, and the residual convolutional block was used to deepen the depth of bottleneck layers. Both the two blocks were employed to enhance the feature extraction of U-net. The experiments were based on the MICCAI 2017’s WMH Challenge datasets, and the results showed that the Dice similarity coefficient (DSC) of our method was 0.83 and the F1 score was 0.84, which were higher than those of compared methods. Through visual observation of the segmentation results, our method cans not only accurately segment large lesion areas, but also distinguish small lesions which are difficult to segment for conventional U-net models.