Research on the fusion method of 2D and 3D UNet in pulmonary nodules segmentation task

Abstract

In segmentation task of pulmonary nodules CT image, the U-shaped convolution neural network (UNet) has made great progress in segmentation accuracy and stability. However, for the three-dimensional (3D) connection characteristics of medical sequence slices, 3D segmentation networks often can’t give full play to their potential and advantages constrained by such factors as the number of marked samples and the computing capacity of equipment. For this reason, this paper focuses on the fusion methods of 2D and 3D UNet in medical image segmentation task: the first is to give full play to the advantages of 2D and 3D networks respectively, and propose a segmentation network model based on two-level hierarchical structure; the second is to make full use of the context information between the image slices and propose a 2.5D segmentation network model. The experiments based on LIDC-IDRI dataset show that the above two methods achieve a good balance in terms of computational performance, segmentation accuracy and model stability.