基于支持向量机的颈动脉超声图像内中膜厚度测量

Carotid intima-media thickness measurement in ultrasound image based on support vector machine

为解决人工标定的繁琐、非客观等问题,本文提出一种基于支持向量机的全自动分割算法。该算法采用K-means对图像像素进行聚类,根据聚类结果和聚类中心对图像进行标准化处理,并进行图像分割提取感兴趣区域。根据训练样本训练支持向量得出分类模型,将感兴趣区域的像素分为边界点和非边界点。然后将边界点再次分类为管腔-内膜边界点和中膜-外膜边界点。最后采用启发式搜索对分类结果进行甄别,去除错分类的像素点。本文采用80幅颈动脉超声图像进行实验,比较实验结果与金标准,内中膜厚度平均绝对误差为(46.08±23.50)μm,平均每幅图像处理时间为0.88 s。实验结果表明全自动分割算法具有快速、全自动等特点,测量结果与金标准具有较高的一致性,能满足临床应用的实际要求。

A fully automatic segmentation（AS） algorithm for the intima-media thickness（IMT） measurement was proposed in the paper to solve the drawbacks of traditional measurement, such as cumbersome manual tracing and non-objectivity. The image pixels were clustered by using K-means of the proposed algorithm. Based on the cluster results and cluster center, the image was normalized and the region of interest（ROI） was extracted by image segmentation. Support vector machines（SVM） was trained by training samples to classify the pixels of ROI into IMT boundary pixels and non-IMT boundary pixels,and the IMT boundary pixels were classified into lumen-intima interface pixels and media-adventitia interface pixels. A heuristic searching method of column-by-column were applied to debug the classification result. A set of 80 ultrasound images of common carotid artery were used to test the proposed method. Comparing experimental results with the ground truth, the mean absolute error of IMT was（46.08±23.50） μm, and the average processing time of each image was 0.88 s. The experience shows the measured results of fully AS algorithm has a high consistency with ground truth, and fully AS algorithm meets the clinical requirements, with advantages of high efficiency and automation.