A novel end-to-end brain tumor segmentation method using improved fully convolutional networks.

Accurate brain magnetic resonance imaging (MRI) tumor segmentation continues to be an active research topic in medical image analysis since it provides doctors with meaningful and reliable quantitative information in diagnosing and monitoring neurological diseases. Successful deep learning-based proposals have been designed, and most of them are built upon image patches. In this paper, a novel end-to-end brain tumor segmentation method is developed using an improved fully convolutional network by modifying the U-Net architecture. In our network, an innovative structure referred to as an up skip connection is first proposed between the encoding path and decoding path to enhance information flow. Moreover, an inception module is adopted in each block to help our network learn richer representations, and an efficient cascade training strategy is introduced to segment brain tumor subregions sequentially. In contrast to those patchwise methods, our model can automatically generate segmentation maps slice by slice. We have validated our proposal by using imaging data from the Multimodal Brain Tumor Image Segmentation Challenge (BRATS) 2015 and BRATS 2016. Experimental results compared with U-Net suggest that our method is 2.6%, 3.9%, and 5.2% higher (by using the BRATS 2015 training dataset) as well as 2.8%, 3.7%, and 8.1% (by using the BRATS 2017 training dataset) higher in terms of complete, core and enhancing tumor regions, respectively. Quantitative and visual evaluation of our method has revealed the effectiveness of the proposed improvements and indicated that our end-to-end segmentation method can achieve a performance that can compete with state-of-the-art brain tumor segmentation approaches.