Abstract
Manual segmentation during clinical diagnosis, is considered as time-consuming and depend to the neuroradiologists level of expertise, however due to the large spatial and structural variability of brain tumors in shapes and sizes besides to the tumor sub-region voxels'high in-homogeneity could make a reliable and accurate and automated segmentation a challenging task. We proposed in this paper, an efficient and fully automatic deep-learning approach for Gliomas 'brain tumor segmentation in multi-sequences Magnetic Resonance imaging (MRI). The proposed method is an optimization on the U-Net based on Fully Convolutional Networks (FCNs) called 'U-Net DFCN' in which we introduced the fusion of multiple MRI modalities to incorporate features from different scales, furthermore, to address the problem of data heterogeneity due to difference in acquisition algorithms and MRI scanner technologies, we proposed an intensity normalization followed by data augmentation techniques in the preprocessing step which though not conventional (usual) in deep FCN-based segmentation approaches. Our method was evaluated on the Multimodal Brain Tumor Image Segmentation (BRATS 2018) training and validation datasets, experimental resulted showed the good performance of the proposed approach outperforming several recent state-of-the-art segmentation methods, achieving a Dice score Coefficient (DSC) of 0.88, 0.87 and 0.81 for complete tumor, tumor-core and enhancing-tumor respectively.