序列血管内超声图像的管腔内膜自动分割算法

针对序列血管内超声(IVUS)图像相邻两帧具有很大相似性的特点，提出一种基于 序列 IVUS 图像配准的自动提取管腔内膜的方法。首先利用形态学运算以及连通分量法提取初 始帧的大致管腔内膜边缘，然后采用前景和背景像素的颜色直方图特征对相邻两帧图像进行建 模并进行配准：利用巴氏系数度量相邻两帧图像之间的相似性，建立仿射变换模型，优化模型 并计算模型参数，从而将轮廓相对准确地定位到管腔内膜附近；最后通过变分法、最速上升法 使目标轮廓曲线准确地收敛到管腔内膜处。以经验医师提取的管腔内膜作为评价标准，分别 与文献17]和4]相比，该方法的均方根误差分别平均减少了 0.124 和 0.063，相对差异度上分 别平均减少了 0.51%和 0.16%。实验结果证明，该方法可以准确地提取到序列 IVUS 图像的管 腔内膜。

Automatic Segmentation Algorithm of Lumen in Sequence IVUS Images

For the sequence intravascular ultrasound (IVUS) images, there are great similarities between two adjacent frames, and a new method based on sequential IVUS image registration for automatic extraction of lumen is proposed. Firstly, in order to realize the automation of extraction process, the morphological operation and the connected component method are used to extract the approximate lumen contour of the initial frame. Secondly, the target region, which includes both foreground and background pixels, is implicitly represented by a level set. The color histogram feature of target region foreground pixels and background pixels are modeled on two adjacent frames because of its merits such as independence of scaling and rotation, robustness to partial occlusions, low computational cost, etc. The Bhattacharyya coefficient measures the similarity between two adjacent frames. The higher the Bhattacharyya coefficient between target model and candidate target model is, the higher the similarity between them is. By establishing affine transformation model, the contour is more accurately located near the lumen. Finally, the segmentation procedure refines the affine transformation estimated in the registration stage, and computes the target’s true shape accurately. The target contour curve is accurately converged to the lumen of the IVUS image by the variational method and the steepest ascent method. Compared with the literature 17], the RMSE of this method is reduced by 0.124 on average, and the RDD is reduced by 0.51% on average. Compared with literature 4], the RMSE of this method is reduced by 0.063 on average, and the RDD is reduced on average by 0.16%. The experimental results show that the method can accurately extract the lumen of several consecutive frame IVUS images.