Robust brain tumor segmentation with deep residual supervision and mixed precision training

Abstract

Segmentation of brain tumors from MRI data is an application of great clinical importance in diagnostic evaluation, treatment and operational planning processes. In recently proposed deep learning techniques, supervision is commonly applied to network output segmentation maps, which may lead to deficiencies in learning features in early network stages. In addition, early termination of training or restricting the number of model parameters in order to limit the computational load caused by three-dimensional architectures that process volumetric MRI data may cause performance losses. The novel segmentation method proposed in this study enhanced sensitivity to information in MR images by applying deep residual supervision on feature maps in decoder stages of the neural network. Additionally, it reduces computational complexity by using mixed precision training algorithms, thus providing effective training in short run times. Experiments on the BraTS dataset show that the proposed model yields higher performance than reference techniques while improving computational efficiency.