基于多模态卷积神经网络的脑血管提取方法研究  

Research on Brain Vessel Extraction via Multi-Modal Convolutional Neural Networks

作　　者：秦志光[1] 陈浩[1] 丁熠[1] 蓝天[1] 陈圆[1] 沈广宇[1] 

机构地区：[1]电子科技大学信息与软件工程学院,成都610054

出　　处：《电子科技大学学报》2016年第4期573-581,共9页Journal of University of Electronic Science and Technology of China

基　　金：国家自然科学基金广东联合基金(U1401257);国家自然科学基金(61300090;61133016);四川省科技厅应用基础计划(2014JY0172);广东省电子信息产品可靠性技术重点实验室开放基金(2013A061401003)

摘　　要：提出了一种基于多模态的卷积神经网络对脑部CT血管造影图像(CTA)进行分割,从而实现脑血管的单独提取。该方法首先对原始CTA图像进行高斯和拉普拉斯处理,并将处理后的图像与原始图像共同构成多模态图像作为输入,然后通过多个并行的卷积神经网络对多模态图像进行分割,最终将所有的分割结果通过线性回归进行融合从而提取出脑血管。该文通过一系列的实验不仅证明了卷积神经网络在脑血管分割上的有效性,而且证明了本文所提出方法的分割效果比现有的脑血管分割算法更加出色。This paper presents a method based on multi-modal convolution neural networks to segment the brain CT angiography image （CTA） for brain vessel. This method firstly processes the original image by adopting the Gaussian and Laplacian filter, respectively, and constructs the multi model image as the input by combining processed images with the original image together. Next, these multi model images will be respectively segmented through a number of parallel convolutional neural networks. Finally, all the segmentation results are fused by employing the linear regression to extract the brain vessel. By evaluating the experiment with the real data acquired from the hospital, it can be proved that the convolution neural network is an effective method for segmenting the cerebral vessels. Moreover, the final result shows that our proposed segmenting method is more accurate than the existing algorithms.

关 键 词：脑血管分割 血管造影图像 卷积神经网络 多模态 

分 类 号：TP315.69[自动化与计算机技术—计算机软件与理论] TP181[自动化与计算机技术—计算机科学与技术]