一种用于超声图像序列分割的水平集演化方法

A Level Set Evolution Method for Ultrasound Image Sequence Segmentation

超声图像具有低信噪比、边界模糊、边界部分缺失、灰度不均等特点,对它的分割极具挑战性.而图像分割又是图像定量、定性分析的关键环节,分割的精确性对后续的分析、处理工作影响重大.距离保持水平集演化(DRLSE)方法对超声图像中出现的弱边界、被部分遮挡边界的分割较差,容易受噪声和灰度不均的影响,因此易造成弱边界泄漏、局部最优等误分割问题;并且初始轮廓对位置敏感,这使得分割的正确性严重依赖初始轮廓位置的选择,故不能对图像进行批量处理.为此提出了一种优化策略:融合基于局部区域的灰度信息和基于边缘的梯度信息构造新的边缘停止函数和面积项权系数,使得演化曲线不仅能够自适应地改变演化方向更有利于对图像序列的处理,同时对斑点噪声和灰度不均问题也有很好的抑制能力;另外,构造了一个先验形状约束项,利用前一帧的分割结果对当前帧的分割进行约束,促进曲线正确演化至目标边界,使得对边界部分遮挡的图像也有着更精确的分割效果.通过合成图像和真实超声图像对分割算法进行了性能分析,设计了基于边缘的豪斯多夫距离和平均绝对距离对算法分割轮廓和医生分割轮廓之间的距离差异性进行度量,实验证明优化策略相比于DRLSE模型和其传统优化模型,有着更高的分割精度,分割效果更出色.

Due to low signal-noise ratio,blurry boundaries,partially occluded boundaries,and intensity inhomogeneity,ultrasound image segmentation is quite challenging.Image segmentation plays a vital role in quantitative and qualitative analysis of ultrasonic images,and the accuracy of segmentation has a great influence on the subsequent processing work.Distance regularized level set evolution(DRLSE)has a poor segmentation effect on the weak boundary and partially occluded boundary appearing in the ultrasonic image,and it is very sensitive to image noise and intensity inhomogeneity.It is therefore easy to cause weak boundary leakage,local optimum,and other missegmentation.In addition,the DRLSE is highly dependent on initial contour positions,and hence,the image cannot be batch processed.In view of the above defects,a new edge-stop function and weighting coefficient-of-area term have been defined.Based on local grey scale and edge gradient,this new function and term have been devised so that the evolution curve could not only adaptively change the evolution direction,but also facilitate the processing of image sequences,while effectively suppressing noise and intensity inhomogeneity.A prior shape constraint was constructed to constrain the segmentation of the current frame via the segmentation result of the previous frame.The curve was correctly evolved to the target boundary so that the image partially-occluded boundary also displayed a more precise segmentation effect.The performance of the segmentation algorithm was analyzed by synthetic image and real ultrasound image.The edge-based Hausdorff distance and the mean absolute distance were designed to measure the distance difference between the algorithm’s segmentation contour and the doctor’s segmentation contour.The experimental results showed that,compared with DRLSE and some optimization models of DRLSE,the proposed method improves both segmentation accuracy and effect.