基于高斯混合模型和核密度估计的全身骨骼SPECT图像分割算法研究

A Research on the Segmentation Algorithm for the Whole Body SPECT Image via the Gaussian Mixture Model with Kernel Density Estimation

目的提出一种基于高斯混合模型的骨扫描图像分割算法,可自动识别全身骨骼SPECT图像中的病变区域。方法首先对二维全身骨骼SPECT图像进行锐化、平滑、灰度变换等预处理;然后采用核密度估计方法拟合出预处理图像的像素概率密度函数曲线,并根据曲线的峰值点确定期望最大值(EM)算法的初始值;再应用高斯混合模型对图像进行分割;最后使用模板匹配算法排除误识别的区域。结果应用本研究提出的图像分割算法所得到的图像清晰度和对比度优于其他图像分割算法,且本研究提出的图像分割算法的相似性测度明显高于其他图像分割算法,平均耗时最短。结论基于高斯混合模型和核密度估计的全身骨骼SPECT图像分割算法是一种高效、实用的骨扫描图像分割算法。

Objective To propose a novel segmentation algorithm for the whole-body bone scan image based on the Gaussian mixture model（GMM） which is used for the automatic recognition of the lesion area. Methods First, we sharpened and smoothed the 2D SPECT whole-body scan image for preprocessing. Second, Gaussian kernel density estimation was adapted to obtain the initial value of the expectation-maximization（EM） algorithm by fitting the curve of probability density function. Then we segmented the image using the GMM algorithm. Finally, the template match method was used to eliminate the wrong recognized areas. Results From subjective evaluation, the presented segmentation method can provide clearer and more detailed activity structures and improve the image quality. Quantitatively experimental results indicate that the GMM algorithm can generate a higher degree of Tanimoto similarity than other methods, and has a less running time. Conclusion Kernel density estimation can effectively prevent the blindness of the initial value selection in the EM algorithm. Thus the lesion areas will be segmented accurately by combination of the GMM and EM method. Therefore, the proposed method is a feasible algorithm for the whole-body bone scan image segmentation.