A cascaded 2.5-dimensional (2.5D) convolutional neural network is proposed. The task is divided into three sub-tasks of whole tumor segmentation, tumor core segmentation and enhancing tumor segmentation, and the results are combined to generate the final result. In each sub-task, the three-dimensional (3D) images are horizontally, coronally and sagittally cropped to generate 2.5D images. The 2.5D images are fed into the proposed 2.5D V-Net for training. The 2.5D segmentation results are concatenated as the 3D results to generate the segmentation results of different sub-tasks. The results show that the average dice scores for the segmentation of whole tumor, tumor core and enhancing tumor by the proposed method are 0.9071, 0.8542, and 0.8140, respectively, which basically meet the clinic need.