Fully convolutional deep residual neural networks for brain tumor segmentation

Abstract

In this paper, a fully convolutional residual neural network (FCR-NN) based on linear identity mappings is implemented for medical image segmentation, employed here in the setting of brain tumors. Inspired by deep residual networks which won the ImageNet ILSVRC 2015 classification challenge, the FCR-NN combines optimization gains from residual identity mappings with a fully convolutional architecture for image segmentation that efficiently accounts for both low- and highlevel image features. After training two separate networks, one for the task of whole tumor segmentation and a second for tissue sub-region segmentation, the serial FCR-NN architecture exceeds state-of-the art with complete tumor, core tumor and enhancing tumor validation Dice scores of 0.87, 0.81 and 0.72 respectively. Despite each FCR-NN comprising a complex 22 layer architecture, the fully convolutional design allows for complete segmentation of a tumor volume within 2 s.