Brain tumor segmentation is considered as a challenging task in the field of medical image analysis. Currently many attempts based on deep neural networks have provided a promising solution, but they generally implement segmentation by capturing the context information on either the global or regional level. In this paper, we propose a localization segmentation network (Loc-Net) that can capture both the global context information, whilst simultaneously focus on the tumor region. Specifically, the proposed framework contains three pathways: Localization pathway, Mask pathway, and Segmentation pathway. Localization pathway intends to localize the tumor regions and determine their bounding boxes. Mask pathway is then designed to transform the bounding boxes as mask features by the residual network, so that the later convolution operations can focus on the content inside the bounding box. Segmentation pathway uses dilate convolution to capture the global context. These features are fused with masked features generated in the mask pathway to get the final prediction. Furthermore, we use the ResNet model to extract comprehensive features from the provided image data, the model is also revised in our works by fusing shallow features with deep features in a different resolution to promote its representation capacity. Experiments on two brain tumor segmentation datasets demonstrate the validity of the proposed three-pathway strategies, and also show that our model can achieve state-of-the-art segmentation accuracy.