A Compound Neural Network for Brain Tumor Segmentation
Hang Zhao, Yucui Guo, Yujing Zheng · 2019
Brain tumor image segmentation has been a time-consuming and laborious task for decades. Since deep learning has been applied to the image field, there have been many efforts to find automatic segmentation methods. This paper proposes an end-to-end image segmentation system based on neural network algorithms. First we develop a novel Convolutional Neural Network architecture named Hybrid Two-path Convolution that combines coarse and fine features obtained by different paths. Besides, traditional method for up-sampling often leads to insufficient utilization of network parameters. By endues geometric meaning to different channels, our Region-based Unpooling Convolution can effectively improve the detailed feature capture capabilities of the network without wasting any parameters. Due to the imbalance between labels, we design category based loss function as a solution. Experiments on public datasets demonstrate that the proposed system is competitive with the state-of-the-art methods.