基于U-Dense-net网络的DSA图像冠状动脉血管分割  

Coronary artery segmentation of DSA images based on U-Dense-net network

作　　者：王卓英 童基均[1] 蒋路茸 潘哲毅 WANG Zhuoying;TONG Jijun;JIANG Lurong;PAN Zheyi(School of Information Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China;Information Office,Hospital of Coast Guard Corps of Chinese People′s Armed Police Forces,Jiaxing 314000,China)

机构地区：[1]浙江理工大学信息学院,杭州310018 [2]武警海警总队医院信息科,浙江嘉兴314000

出　　处：《浙江理工大学学报（自然科学版）》2021年第3期390-399,共10页Journal of Zhejiang Sci-Tech University(Natural Sciences)

基　　金：国家自然科学基金项目(61602417);浙江理工大学基础研究项目(2019Q042);浙江理工大学“521人才培养计划”。

摘　　要：冠状动脉血管是研究心血管疾病的重要基础,为准确分割DSA(Digital subtraction angiography)图像冠状动脉血管,提高训练过程中血管特征的有效利用率,提出了一种基于U-Dense-net网络的分割方法。该方法首先对数据集进行限制对比度直方图均衡化预处理;然后对预处理结果进行图像粗分割,基于U-Dense-net网络,在解码器部分融合密集残差块和注意力机制实现深度神经网络模型,加强特征映射,充分提取局部特征,实现血管与背景的分类;最后利用形态学处理、阈值分割、基于多点区域生长的连通域分析进行图像细分割,实现血管的提取。将测试结果和3位专家手工标注的标准图进行对比分析,结果表明:该数据集的分割结果精确率、召回率、F1分数分别为83.22%、89.81%、86.04%,3种特性曲线下的平均面积为0.9923。与其他方法比较,该方法提取到的血管信息较为完整,为精确分割冠状动脉血管提供了一种解决方案。Coronary artery is an important basis for the study of cardiovascular diseases. In order to segment coronary arteries accurately in DSA(digital subtraction angiography) image and improve the effective utilization of vascular features in the training process, this paper proposes a segmentation method based on U-Dense-net. Firstly, the data set is preprocessed by limiting contrast and equalizing histograms. Secondly, the preprocessed results are roughly segmented. Based on U-Dense-net, our study implements a deep neural network model by fusing the dense residual block and attention mechanism in the decoder to strengthen feature mapping, fully extract local features and realize the classification of blood vessels and backgrounds. In the end, morphological processing, threshold segmentation and connected component analysis based on multi-point region growth are used to segment the images finely and extract vessels. The test results are compared with the standard drawings annotated by three experts manually, and it is found that the accuracy rate, recall rate and F1 score of the segmentation result of this data set are 83.22%, 89.81% and 86.04% respectively, and the mean area under three ROC curves is 0.9923. The comparison with other methods shows that the proposed method can extract complete blood vessel information and provide an effective solution for the accurate segmentation of coronary arteries.

关 键 词：冠状动脉血管 图像分割 U-Dense-net 密集残差块 注意力机制 深度神经网络 DSA 

分 类 号：TS391.4[轻工技术与工程]