Deep learning has become the method of choice for tumor segmentation. Most deep learning algorithms incorporate a multi-modal approach, as different MR modalities are optimized to detect different aspects of tumor. However, modalities are often missing or unusable due to artifacts. In such cases, it is difficult to perform robust automatic tumor segmentation. We demonstrate that a convolutional neural network can be used to synthesize FLAIR MR images that have high similarity with real FLAIR images. Furthermore, we show that the use of these synthetic images can improve segmentation performance.