Abstract: Accurate segmentation of multiple gliomas from multimodal MRI is a prerequisite for many precision medical procedures. To effectively use the characteristics of glioma MRI and im-prove the segmentation accuracy, we proposes a multi- Dice loss function structure and used pre-experiments to select the good hyperparameters (i.e. data dimension, image fusion step, and the implementation of loss function) to construct a 3D full convolution DenseNet- based image feature learning network. This study included 274 segmented training sets of glioma MRI and 110 test sets without segmentation. After grayscale normalization of the image, the 3D image block was extracted as a network input, and the network output used the image block fusion method to obtain the final segmentation result. The proposed structure improved the accuracy of glioma segmentation compared to a general structure. In the on-line assessment of the open BraTS2015 data set, the Dice values for the entire tumor area, tumor core area, and enhanced tumor area were 0.85, 0.71, and 0.63, respectively.