文献类型：期刊文献

英文题名：Dual-pathway DenseNetswith fully lateral connections for multimodal brain tumor segmentation

作者：Hu, Jingyu[1];Gu, Xiaojing[1];Gu, Xingsheng[1]

机构：[1]East China Univ Sci & Technol, Key Lab Adv Control & Optimizat Chem Proc, Minist Educ, Shanghai, Peoples R China

年份：2021

卷号：31

期号：1

起止页码：364

外文期刊名：INTERNATIONAL JOURNAL OF IMAGING SYSTEMS AND TECHNOLOGY

收录：;EI(收录号：20203509113056);WOS:【SCI-EXPANDED(收录号:WOS:000563868500001)】；

基金：National Natural Science Foundation of China, Grant/Award Numbers: 61973122, 61973120, 61573144

语种：英文

外文关键词：brain tumor; convolutional network; dense network; multimodalities; segmentation

摘要：Multimodal medical image data provide different structured and functional information, which helps segment brain tumor and gets a reliable and accurate diagnosis. Segmenting brain tumors in magnetic resonance imaging (MRI) is a challenging task because brain tumors can be at any location with changeable shape and size. Existing deep neural networks for brain tumor segmentation use few connections to fuse multilevel information. To make use of multilevel information from multimodal MRIs, we propose dual-pathway DenseNets with fully lateral connections (DP-DenseNets), a three-dimensional (3D) fully convolutional neural network that uses dense connectivity to construct dual-pathway architecture to multimodal brain tumor segmentation problem. Each two similar imaging modalities have a pathway, for one thing, the bottom-up pathway with dense connectivity is developed for extracting features; another, the top-down pathway concatenates the features of the bottom-up pathway in all layers. Dual pathways with different loss functions and fully lateral connectivity from the bottom-up pathway to the top-down pathway provide an abundant combination of different levels of features. Comparing to these fusion schemes such as input-level fusion and later-level fusion, this architecture leverages semantics from low to high levels, which is provided by fully lateral connectivity. Our model is evaluated on the dataset from Brain Tumor Segmentation Challenge 2017 (BRATS 2017), and the experiments show that our method achieves better performance than other 3D networks.