Multi-task Fully Convolutional Network for Brain Tumour Segmentation

In this paper, a novel, multi-task fully convolutional network (FCN) architecture is proposed for automatic segmentation of brain tumour. The proposed network builds on the hierarchical relationship between tumour substructures with branch and leaf losses imposed and optimised simultaneously. The network takes multimodal MR images along with their symmetric-difference images as input and extracts multi-level contextual information, firstly by the branch losses which are then fed to the leaf loss in a combination stage. The model was evaluated on BRATS13 and BRATS15 datasets and results show that the proposed multi-task FCN outperforms single-task FCN on all sub-tasks. The method is among the most accurate available and its computational cost is relatively low at test time.