3D glioma magnetic resonance imaging has different tumor shapes and blurred edges. The segmentation method based on 2D Convolutional Neural Network cannot segment the three-dimensional image well. In order to accurately segment the tumor in the three-dimensional image, a 3D Convolutional Neural Network brain tumor image segmentation method fused with multi-scale feature information is proposed. The feature information is extracted by parallel 3D dilated convolution, and the information of different receptive fields is fused. The Dice loss and the BCE loss are combined to form a new loss function and cooperate with the identity mapping to further improve the segmentation accuracy. The model was verified on the BraTs2020 data set. The Dice coefficients of the whole tumor area, core area, and enhancement area segmented by the model are 89.1%, 83.9%, 82.6%. The model was verified on the LGG brain tumor image data set, and the Dice coefficient reached 93.3%. The segmentation method can not only accurately segment three-dimensional glioma images, but is also suitable for segmentation of two-dimensional glioma images.