This paper presents a fully automatic method for brain tumor segmentation using Deep Neural Networks (DNNs). The method is specifically designed to segment glioblastomas, both low and high grade, from Magnetic Resonance Imaging (MRI) data. The authors explore various Convolutional Neural Network (CNN) architectures tailored to image data, focusing on models that can exploit both local and global contextual features. A novel CNN architecture is introduced, which uses a convolutional implementation of a fully connected layer to achieve a 40-fold speedup. The method also includes a two-phase training procedure to address the imbalance in tumor labels and a cascaded architecture to improve segmentation accuracy. The results, reported on the 2013 BRATS test dataset, show that the proposed method outperforms state-of-the-art methods while being over 30 times faster. The contributions of the work include a fully automatic method ranked second on the BRATS 2013 scoreboard, a novel two-pathway architecture, and a cascaded architecture that efficiently models label dependencies.This paper presents a fully automatic method for brain tumor segmentation using Deep Neural Networks (DNNs). The method is specifically designed to segment glioblastomas, both low and high grade, from Magnetic Resonance Imaging (MRI) data. The authors explore various Convolutional Neural Network (CNN) architectures tailored to image data, focusing on models that can exploit both local and global contextual features. A novel CNN architecture is introduced, which uses a convolutional implementation of a fully connected layer to achieve a 40-fold speedup. The method also includes a two-phase training procedure to address the imbalance in tumor labels and a cascaded architecture to improve segmentation accuracy. The results, reported on the 2013 BRATS test dataset, show that the proposed method outperforms state-of-the-art methods while being over 30 times faster. The contributions of the work include a fully automatic method ranked second on the BRATS 2013 scoreboard, a novel two-pathway architecture, and a cascaded architecture that efficiently models label dependencies.