Automatic segmentation of brain tumor images is difficult to achieve owing to the diversity in tumor shapes and severe imbalance in the segmentation categories. Conventional convolutional neural network can hardly predict high precision segmentation images. To solve the abovementioned problems, an improved model based on the original three-dimensional (3D)-Unet was proposed, which replaced the conventional convolution module with a hybrid dilated convolution module to exponentially increase the receptive field of neurons, reducing the network depth and avoiding scenarios wherein small targets could not be recovered during up-sampling. Furthermore, the hybrid loss function was used to replace the original Dice loss function to increase the penalty faced by the model when classification errors of sparse classes occurred, forcing the model to learn the features of these classes better. Experiment results showed that the hybrid dilated convolution module and the hybrid loss function could respectively improve the prediction accuracy of the whole tumor region and the core tumor region. Multiple performance parameters of brain tumor automatic segmentation were improved using this 3D-HDC-Unet model.