Despite remarkable progress, 3D whole brain segmentation of structural magnetic resonance imaging (MRI) into a large number of regions (>100) is still difficult due to the lack of annotated data and the limitation of GPU memory. To address these challenges, we propose a semi-supervised segmentation method based on deep neural networks to exploit the plenty of unlabeled data by extending the self-training method, and improve the U-Net model by designing a novel self-ensemble architecture and a random patch-size training strategy. Further, to reduce the model storage and computational cost, we get a compact model by knowledge distillation. Extensive experiments conducted on the MICCAI 2012 dataset demonstrate that our method dramati- cally outperforms previous methods and has achieved the state-of-the-art per- formance. Our compact model segments an MRI image within 3 s on a TITAN X GPU, which is much faster than multi-atlas based methods and previous deep learning methods.