We proposed a deep learning approach for automatic segmentation of three-dimensional gliomas magnetic resonance images(MRI). First, we used a three-stage cascaded strategy to sequentially segment the subregions of gliomas. Second, to further improve the segmentation accuracy, we used intraslice and interslice convolutions, introduced additional multi-levels feature fusing, and implemented dilated convolution. Third, to produce a fine-grained output, conditional random fields as recurrent neural network were adopted as a part of network structure. Finally, we combined two types of loss functions in the training procedure to further improve the segmentation accuracy. We applied our method on the BraTS 2018 dataset and achieved a Dice score of 0.9093, 0.8254, and 0.7855 and the Hausdorff distance of 3.8188, 7.8487, and 4.3264 for the whole tumor, tumor core, and enhanced tumor, respectively. The proposed methods achieved better performance than most brain tumor segmentation methods.