Brain Tumor Automatic Segmentation Using Fully Convolutional Networks

Brain tumor segmentation has the most important significance in tumor diagnose, analysis, and treatment. Magnetic resonance imaging (MRI) provides rich and valuable information for physicians to analyze tumors. However, intracranial tumors can grow anywhere in the brain and have complexed structure and amorphous shape, which makes manual segmentation inconceivable to mark millions of imaging data from MRI. This paper explores a new method for brain tumor automatic segmentation of MRI images using the Fully Convolutional Networks. We start the intensity normalization in pre-processing phase to improve comparability of different scans. Then, we take the whole slices across four modalities as four-channel input images to train network, which can overcome the individual difference of brain tumor and difficulty to set patch size in the methods based on the center pixel classification of image patch. The experiment results of 220 patients from Brain Tumor Segmentation Challenge 2015 database show that the proposed method is effective. The average Dice Similarity Coefficient metric are 84.21%, 77.36%, and 85.07% in the complete, core, and enhanced regions, respectively. The average time of segmenting one slice is around 0.15 s. The proposed method can adapt to the differences of brain tumors, quickly accurately.