多尺度非参数化水平集的超声心动图分割

Multiscale Non-Parametric Level Set Segmentation of Ultrasound Echocardiography

针对超声心动图噪声很大、提取目标区域边界不够平滑完整的问题,将非参数技术与水平集相结合,提出了多尺度非参数化的水平集图像分割方法。利用非局域均值滤波建立尺度空间,保护图像特征,在粗尺度预分割,然后在细尺度优化分割。采用Parzen窗技术对超声心动图的亮度分布进行统计建模,不需要先验假设,引入到水平集框架中,设计了非参数化水平集分割模型。分割实验证明:预分割结果和真实边界的平均绝对距离为2.162,优化后为0.710。该方法可以精确地自动提取感兴趣区域,在图像分割鲁棒性和精确性方面优于常规分割方法。

To solve difficulty that the boundary of segmented objective region is not enough smooth and complete due to ultrasound echocardiography with serious noise,a multiscale segmentation approach combined non-parametric technique with level set,is presented.The nonlocal means filtering（NLM） is performed to create scale space and preserve image features.Pre-segmentation is firstly carried out in a coarser scale image,then an optimized segmentation in a finer scale image,the intensity distribution of the ultrasound echo images is modeled by Parzen window technique without prior assumption.A non-parametric model on level set framework is designed to segment the ultrasound echocardiography.The segmentation experiments show that the mean absolute distance（MAD） between real boundary and pre-segmentation result gets to 2.162,but,and the optimized result only 0.710.The approach outperforms the conventional segmentation methods by accurately and automatically extracting the regions of interest.