Deep Hourglass for Brain Tumor Segmentation

The segmentation of a brain tumour in an MRI scan is a challenging task, in this paper we present our results for this problem via the BraTS 2018 challenge, consisting of 210 high grade glioma (HGG) and 75 low grade glioma (LGG) volumes for training. We train and evaluate a convolutional neural network (CNN) encoder-decoder network based on a singular hourglass structure. The hourglass network is able to classify the whole tumour (WT), enhancing (ET) tumour and core tumour (TC) in one pass. We apply a small amount of preprocessing to the data before feeding it to the network but no post processing. We apply our method to two different unseen sets of volumes containing 66 and 191 volumes. We achieve an overall Dice coefficient of 92% on the training set. On the first unseen set our network achieves Dice coefficients of 0.66, 0.82 and 0.72 for ET, WT and TC. On the second unseen set our network achieves Dice coefficients of 0.62, 0.79 and 0.65 on ET, WT and TC.