Deep learning for segmentation of multi-modal Magnetic Resonance Images from brain tumor

Abstract

Brain tumor segmentation plays a critical role in both tumor diagnosis and treatment. Therefore, BraTS challenge encourages researches to improve the automatic brain tumor segmentation methods, which will lessen the burden on medical professionals. Currently, most state-of-the-art solutions are based on deep learning, because of the availability of advanced GPUs and labelled data. In this project, we will explore 3D encoder-decoder based networks for automatic segmentation of brain tumor and intra-tumor structures using MRI images, and we will attempt to improve the performance further. In our model we extracted 3D patches to predict the class labels for all pixels for each patch, this will help us to overcome the class imbalance problem and to reduce the memory demand. We also applied augmentation techniques to minimise the risk of overfitting by increasing the size of the training data. Our networks were implemented in Keras, with a TensorFlow backend. Finally, we have evaluated our proposed network on the test dataset provided by the Brain Tumor Segmentation 2018 (BraTS 2018) Challenge. Correspondingly, the Dice similarity index of 0.86, 0.78 and 0.76 for the segmentation of the whole tumor, tumor core and enhanced tumor, respectively.