改进分水岭与K-means++结合的人体脊柱图像分割方法

Human Spine Image Segmentation Method Based on Improved Watershed and K-means++

由于在人体脊柱图像分割中,分水岭算法存在过分割现象以及对微弱边缘、噪声非常敏感的问题,故对其进行改进。原始人体脊柱CT图像存在许多不必要的局部极小值,首先利用K-means++聚类算法进行区域分类,减少错误的局部极小值;然后利用形态学图像处理技术对初始分割图像进行去噪处理,使图像变得平滑;接下来提取区域最大值标记为图像前景,将阈值分割得到的图像标记为背景;最后通过分水岭变换得到人体脊柱分割结果图。实验结果表明,该算法能实现对人体脊柱图像的准确分割,其Dice系数、Jaccard系数与Precision系数均值分别为89.2%、82.3%和85.4%,相比当前主流算法分别提高了15.2%、12.7%与10.3%。

Due to the over segmentation phenomenon of watershed algorithm in human spine image segmentation and the problem of being very sensitive to weak edges and noise,this paper improves it.There are many unnecessary local minima in the original human spine CT imag⁃es.Then,morphological image processing technology is used to denoise the initial segmented image to make the image smooth;The maximum value of the extracted region is marked as the foreground of the image,and the image obtained by threshold segmentation is marked as the background.Finally,the segmentation result of human spine is obtained by watershed transform.Experiments show that the algorithm can achieve accurate segmentation of human spine image.The mean values of Dice coefficient,Jaccard coefficient and precision coefficient are 89.2%,82.3%and 85.4%respectively,which are 15.2%,12.7%and 10.3%higher than the current mainstream algorithms.