| 引用本文: | 朱锴,陶攀,付忠良,陈晓清.利用关键点检测算法的超声图像定点测量[J].哈尔滨工业大学学报,2018,50(11):67.DOI:10.11918/j.issn.0367-6234.201711029 |
| ZHU Kai,TAO Pan,FU Zhongliang,CHEN Xiaoqing.Ultrasound image fixed point measurement based on landmark detection method[J].Journal of Harbin Institute of Technology,2018,50(11):67.DOI:10.11918/j.issn.0367-6234.201711029 |
|
| 本文已被：浏览 1419次 下载 1303次 |
码上扫一扫！ |
|
|
| 利用关键点检测算法的超声图像定点测量 |
|
朱锴1,2,陶攀1,2,付忠良1,陈晓清1,2
|
|
(1．中国科学院 成都计算机应用研究所, 成都 610041;2. 中国科学院大学,北京 100049)
|
|
| 摘要: |
| 超声图像的定点距离测量在临床医学上十分重要.由于超声图像的噪声大，边缘模糊，因此关键点自动定位在超声图像中很有挑战性.目前的关键点检测算法通常是针对单个关键点位置进行优化，难以在保证每个关键点检测精度的情况下得到准确的测量距离.为使超声图像中关键点的精度和两个关键点之间的距离更加精确，本文提出一种基于级联卷积神经网络的关键点检测算法，该方法采用两个卷积网络从粗略到精细的对关键点进行定位.首先利用第一个网络回归两个关键点的粗略位置，并将包含这两个关键点的小区域送入第二个网络.然后本文提出一种加入距离修正的损失函数，作为第二个网络的优化目标，在第一个网络输出结果的基础上定位最终的关键点位置.实验结果表明，本文提出的级联方法无论是相比传统的级联方式还是回归树方法，本文算法在超声图像的关键点定位上更为精准，并且在最终的距离测量精度上也有很大的提高，在评价标准下比传统级联方法检测精度上提升将近30%.
|
| 关键词: 超声心动图 关键点定位 卷积神经网络 损失函数 自动测量 |
| DOI：10.11918/j.issn.0367-6234.201711029 |
| 分类号:TP391.41 |
| 文献标识码:A |
| 基金项目:四川省科技厅重点研发项目(2017SZ0010); 四川省科技支撑计划项目(2016JZ0035) |
|
| Ultrasound image fixed point measurement based on landmark detection method |
|
ZHU Kai1,2,TAO Pan1,2,FU Zhongliang1,CHEN Xiaoqing1,2
|
|
(1. Chengdu Institute of Computer Applications, Chinese Academy of Sciences, Chengdu,610041, China; 2. University of Chinese Academy of Science, Beijing 100049, China)
|
| Abstract: |
| The measurement by fixed point of ultrasound image is very important in clinical medicine. Because of the large speckle noise and blurred edges in ultrasound images, landmark detection in echocardiography is quite challenging. Meanwhile, current landmark detection algorithm optimizes a single landmark position, and it's difficult to get accurate distance with guaranteeing the accuracy of each landmark. To get more accuracy result of landmark and the distance between two landmarks in ultrasound, cascaded convolution neural network is proposed to automatic detection landmark in echocardiography, our framework adopts two stages of carefully designed deep convolution networks that predict landmark location in a coarse-to-fine manner. Firstly, networks at the first level estimate positions of two landmarks coarsely, the patch includes two landmarks as input of the second stage. Then a loss function with distance correction term is proposed to optimize the second network, and gets the final landmark position based on the output of first stage. Experimental results show that compared to traditional network and regression tree, the proposed method could not only guarantee the accuracy of landmark position but also increase the accuracy of distance, compared with traditional cascaded convolution neural network, and the accuracy of distance is nearly 30% higher than the traditional method.
|
| Key words: echocardiography landmark detection convolution neural network loss function automatic measurement |
