一种普适的基于多尺度滤波和统计学混合模型的血管分割方法

A Universal Vessel Segmentation Method Based on Multi-Scale Filtering and Statistical Mixture Model

血管的精确提取和定位,是实现心脑血管介入手术的关键。多尺度滤波算法可以增强血管目标,同时抑制背景噪声,但并没有把血管从图像背景中区分出来。基于统计学的分割算法,通过对直方图进行拟合实现血管的分类,但需要调整混合模型去拟合特定的图像直方图。为了克服上述问题,提出一种具有固定模型的普适的血管分割方法。首先,利用多尺度滤波算法进行图像预处理。其次,针对滤波增强后数据的直方图曲线,用由3个概率分布函数(1个高斯和2个指数)组成的混合模型进行拟合。期望最大化算法用于混合模型参数的估计。最后,通过最大后验概率分类算法将血管分离出来。为了验证上述方法的有效性,分别在仿真(phantom)数据、磁共振血管造影(MAR)数据和计算机断层血管造影(CTA)数据上进行实验测试。结果表明,所提出的方法在多套仿真数据上的分割误差低于0.3%,同时对于不同模态的血管图像具有很好的分割效果及较强的鲁棒性。

Accurate extraction and localization of blood vessels are the keys to the intervention operation of cardiac and cerebral vessels. Multi-scale filtering strengthens the vessels while weaken the background voxels,but the vessels are still not marked out. Statistical based segmentation method classifies the vessels through model fitting for the histogram curve,but it needs to adjust its model to fit a certain image histogram. To overcome these problems,a universal vessel segmentation method with a fixed model has been proposed in this paper. Firstly,the original image was preprocessed with multi-scale vessel enhancement algorithm. Secondly,a mixture model formed by three probabilistic distributions（ one normal distribution and two exponentials） was built to fit the enhanced data. Expectation maximization algorithm has been used for parameters estimation.Finally,the vessels were segmented by maximum a posteriori classification. To test the effectiveness of the proposed method,experiments have been done on a series of phantoms,magnetic resonance angiography（ MRA） data and computed tomography angiography（ CTA） data. As a result,the segmentation errors of the phantoms are less than 0. 3%. Meanwhile,the proposed method performed well on multi-modality images with strong robustness.