With the rapid increase of the amount of available medical images, computer-based automatic medical image analysis is becoming a promising and prosperous research area, among which, automatic segmentation of medical image is more and more important for its necessity in the clinical diagnosis. In recent years, quite a number of semantic segmentation methods based on deep convolution neural networks (CNNs) have been proposed and shown much better performance than those conventional methods. However, most of them are designed for 2D images based on 2D CNN. In nowadays, more and more medical images are three dimensional. The 3D CNN-based methods are good at dealing with 3D images, but more computationally expensive and require more training samples. Training a 3D CNN model from scratch is much difficult when the training dataset is relatively small. To leverage full information of 3D medical image data and reduce the difficulty of training at the same time, a new end-to-end pseudo-3D fully convolutional densenet model for 3D brain tumor segmentation is presented in this paper. The new method decomposes one 3D convolution into two correlated 2D convolutions to reduce the number of parameters to be tuned. It pre-trains a 2D model on large 2D image datasets first, then adapts it on the relatively small 3D brain tumor dataset to implement 3D brain tumor segmentation. Experiment results show that the new model achieves comparable or even better performance than some well-known networks (e.g. Vnet, 3D-Unet), while reduces the training complexity obviously.