Brain Tumor Segmentation on MR Images Using Anisotropic Deeply Supervised Convolutional Neural Network

Glioma is the most common and cancerous type of early stage brain tumors that arise from glial cells. Deep learning based solutions are flexible in brain tumor image analysis and computer-assisted diagnosis but it does not gain the desired accuracy. Manual brain tumor segmentation is a challenging task for clinicians therefore, researchers are working continuously to improve the accuracy for brain tumor segmentation using automatic segmentation. In this paper we segmented the brain tumor in three regions namely whole tumor, enhancing tumor and non-enhancing tumor using Convolutional Neural Network (CNN) implemented by anisotropic dilated convolution with residual blocks additionally using deeply supervised layers. Experiment shows that, our proposed method achieved the dice scores of 0.91, 0.78, 0.84 for whole tumor, enhancing tumor and non-enhancing tumor respectively, which is better than the BraTS 2017 challenge and other reported approaches.