Brain Tumor Segmentation from MRI Using Pre-segmentation Based on Superpixels and Fully Convolutional Neural Networks

This paper focuses on the development of an effective method for brain tumor segmentation in MR image, which includes two novel approaches. The first approach is an image pre-segmentation algorithm using superpixels for brain tumor segmentation in T1-w, T1-c, T2-w and FLAIR, respectively. We use pre-segmentation as the rough segmentation of the image to get the region of interest (ROI). The images of each subject were already aligned with the T1c and skull stripped. The second segmentation approach is based on Fully Convolutional Network. The performance evaluation for the proposed method is based on MRI data sets of Medical Image Computing and Computer Aided Intervention Society (MICCAI) 2017 Brain Tumor Segmentation Challenge. The numerical results are presented in terms of true positive rate (TPR), positive predictive value (PPV) and Dice coefficient (DICE) for GD-enhancing tumor (ET), tumor core (TC) and whole tumor (WT), respectively, which are the major three components concerned by medical doctor in diagnosis and treatment planning for brain tumor. A comparison of the proposed method and the expert manual method, as well as the other conventional method shows that the proposed method is able to acquire a larger mean value of the Dice similarity coefficient than the other conventional methods do. Therefore, this novel method is an added value for the clinical radiology of brain tumor.