所属单位：计算机科学与技术学院/人工智能学院/软件学院
发表刊物：Lect. Notes Comput. Sci.
摘要：Brain network plays an important role in the diagnosis of brain diseases. Recently, applying convolutional neural networks (CNNs) in brain network has attracted great interests. Although traditional convolution can capture all the nearest neighbors of a point in Euclidean space, it may miss the nearest neighbors of the node in the graph (i.e., in topological space). Hence, how to design a meaningful convolutional operator is a major challenge for brain network classification. Accordingly, in this paper, we propose a node based row-filter convolutional neural network, named NRF-CNN, for brain network classification. The proposed NRF-CNN can learn the high-order representation for brain network without losing important structure information. Specifically, we first introduce the concept of node neighbor within one step. Then, we define a novel row-filter convolutional operator, which can effectively capture local pattern of graph by applying a row scanning on adjacent matrix. Next, we adopt a structure preserved pooling to enrich the node representation by multiply input adjacent matrix and the node representation. Further, we stack several row-filter convolutional layers and structure preserved pooling layers to capture feature representation with more complex information. Finally, we fuse all features learned from each layer by linear weighting. To evaluate the effectiveness of our approach, we compare it with state-of-the-art methods in brain network classification on three real brain network datasets. The experimental results demonstrate that our approach outperforms the others, showing better capacity in capturing meaningful and discriminative representations for brain networks. © Springer Nature Switzerland AG 2018.
ISSN号：0302-9743
是否译文：否
发表时间：2018-01-01
合写作者：Mao, Bingcheng,Huang, Jiashuang
通讯作者：张道强
发表时间：2018-01-01