用于CT心脏分割的几种超像素过分割和块合并的方法比较

Comparison of Several Methods of Superpixel-Based over-Segmentation and Region Merging for Cardiac CT Segmentation

针对从CT图像中提取心脏结构信息还是一个尚未解决的问题,本文利用超像素思想对CT图像进行分割。本文利用4种方法(Ncut算法、熵率、简单线性迭代、均值漂移)进行超像素过分割,并进行了量化比较。进一步通过动态融合方法和谱聚类方法得到分割结果。在动态融合方法中设计了一种相似性度量的计算方法,并对两种合并方法进行了比较。实验表明本文提出的方法用于CT心脏图像的分割是可行的。在四种超像素过分割方法中,简单线性迭代运行速度较快,在各项评价指标中都比较不错。动态融合方法和谱聚类的合并准确性都较高,但谱聚类的运算速度远快于超像素的动态合并。

To tackle the unresolved problem of extracting cardiac structure information from CT images,this paper uses superpixel paradigm to segment CT images.In this paper,four methods(N-cut algorithm,entropy rate,simple linear iterative clustering,mean shift)are used to perform superpixel-based over-segmentation and their quantitative comparison is performed.The segmentation results are obtained by further dynamic fusion method and spectral clustering method.A calculation method of similarity measure is designed in the dynamic fusion method,and the two region merging methods are compared.Experiments show that the proposed method is feasible for segmentation of CT cardiac images.Among the four superpixel-based over-segmentation methods,the simple linear iterative clsutering runs faster and is perfoming well in all evaluation indicators.The accuracy of both dynamic fusion method and spectral clustering is good,but the spectral clustering operation speed is much faster than the dynamic fusion.