结合高斯过程和形变模型的超声图像分割方法

Ultrasound image segmentation method based on gaussian process and deformable model

目标区域的先验形状在基于形变模型的超声图像分割方法中扮演着重要的角色。为了提高先验形状模型对目标轮廓形变细节的建模能力,提出了一种基于高斯过程的统计形状模型。目标的形状被表示成一种离散的随机时间序列;利用高斯过程的性质对训练集中的目标形状变化进行统计学习,从而生成目标的先验形状和先验概率。为给形变模型向目标区域的演化提供观测模型,结合超声图像中目标边缘内外灰度变化特征设计了一种径向纹理特征模型。分割的优化被转化为求最大后验概率的过程。基于真实的临床超声图像实验结果显示,与其他方法相比该方法在复杂形变区域和弱边缘区域提供了更准确和鲁棒的结果。

The shape prior of target region plays an important role in the deformable model based ultrasound image segmen-tation methods. To improve the modeling ability of the shape prior model to the target contour deformation details, a novel Gaussian process based statistical shape model is proposed. Firstly, the target contour is represented as a discrete stochastic time series. And then, the properties of Gaussian process are utilized to statistically learn the variations of the target shapes from training set. In order to provide the observation model for the evolution of deformable model to target region, a radial texture feature model is developed by combining the intensity variation features from the inside to outside of target region. Finally, the segmentation optimization process is cast to compute a maximum posteriori process. The experimental results of real clinical ultrasound image show that, compared with other well known methods, the proposed method provides more accurate and robust results in the complex deformable and weak edge regions.