First, T1-weighted (T1W), T2-weighted (T2W) and T1 enhanced structural MR images of 421 patients were collected, the tumor boundaries of all images were delineated manually by two experienced doctors as the ground truth, the images and ground truth of 346 patients were considered as training set and the remaining images and corresponding ground truth of 75 patients were selected as independent testing set. Second, three single modality, based multi-dimension deep convolutional neural networks (CNN) and three two-modality multi-dimension fusion deep convolutional networks and a multi-modality multi-dimension fusion (MMMDF) deep convolutional neural network were constructed, and the networks were trained and tested, respectively. Finally, the performance of the three methods were evaluated by using three indexes, including Dice, Hausdorff distance (HD) and percentage area difference (PAD). The experimental results show that the MMMDF CNNs can acquire the best performances, followed by the two-modality multi-dimental fusion CNNs, while the single modlity multi-dimension CNNs achieves the worst measures.. This study demonstrates that the MMMDF-CNN combining multi-modality images and incorporating 2D with 3D images features can effectively fulfill accurate segmentation on tumors of MR images from NPC patients.