超声左心耳图像的轮廓自动提取

Automatic Contour Extraction of Left Atrial Appendage from Ultrasound Images

针对经食道超声左心耳图像的分辨率低、对比度低、含有斑点噪声等问题,提出一种结合左心耳解剖位置和超声图像灰度及相位信息的方法,全自动定位常规切片中的左心耳。首先,根据医生采集习惯,以左心耳在标准切面中的解剖位置为先验知识,结合其灰度特性,自动确定分割模型中的初始轮廓;然后,通过线型加权相位和梯度信息构造新的外力项,改进向量场卷积模型,完成左心耳轮廓的自动提取。300张左心耳超声图片测试结果表明,以医生手动勾勒的轮廓作为"金标准",该方法自动提取左心耳的准确性为0.896 9±0.049 4、敏感性为0.905 8±0.076 2、特异性为0.964 5±0.168 7。分割效果优于传统的向量场卷积模型,能够解决自动定位超声图像中左心耳的初始轮廓和弱边界分割的问题。

In order to fulfil the contour extraction of left atrial appendage （LAA） from the low resolution, low contrast and noisy transesophageal ultrasound images automatically, a new method was proposed, which combined the anatomical knowledge of LAA and the information of intensity and phase in ultrasound images. Firstly, to locate the effective initial contour with the characteristic of intensity and geometry from LAA ultrasound images, the anatomy location of LAA in the standard section was taken as the priori-knowledge based on the physicians＇ collecting habit. Then, to improve the convergence performance of classical vector filed convolution （VFC） model, a new external force term was proposed by combing phase and gradient information in the linear weighting method. The test results conducted from 300 frames showed that this method can reach the accuracy of 0. 896 9± 0. 049 4, sensitivity of 0. 905 8 ± 0. 076 2 and specificity of 0. 964 5 ± 0. 168 7 by regarding the contours outlined by physicians as ＂golden standard＂. Comparison with traditional VFC model showed that this method owns better segmentation performance by determining a more suitable initial contour and is more robust to weak edge.