一种层次化的乳腺肿瘤分割方法

Hierarchical segmentation of breast tumor in ultrasound image

超声图像是乳腺癌辅助诊断常用的工具之一.肿瘤分割是乳腺超声图像分析的基础.乳腺超声图像中的灰度不同质性、纹理及形状的多变性等复杂特点使得肿瘤的精确分割较为困难.提出了一种层次化的分割框架.首先将局部灰度聚类假设引入活动轮廓模型作为底层分割模型,对图像进行初始分割;然后提出基于超像素和支持向量机(Support Vector Machine,SVM)的高层分割模型,对初始结果再进行高层分割.在高层分割过程中,首先使用简单线性迭代聚类(Simple Linear Interactive Cluster,SLIC)提取超像素,然后提取超像素的灰度、纹理和局部特征,最后使用SVM进行分类.高层分割模型是基于底层模型的分割结果学习获取的,能够检测到底层模型可能分割错误的区域,与底层模型具有较好的互补性.因此,提出的层次化分割框架具有较好的鲁棒性.在自建乳腺超声数据库上的实验结果证明了提出方法的有效性和鲁棒性.

Breast ultrasound image analysis is commonly used to help breast cancer diagnosis.Breast tumor segmentation is the first step of image analysis and plays an important role for the automatic breast ultrasound image analysis.However,complex characteristics such as intensity inhomogeneity,variance of texture and shape in breast ultrasound images makes accurate tumor segmentation difficult.In order to deal with the complex characteristics appeared in ultrasound images,this paper proposes a hierarchical breast tumor segmentation method.The method is developed with active contour model framework.In the proposed framework,active contour model is first to be used as low-level segmentation model for initial segmentation.Based on the initial segmentation result,a high-level segmentation model based on superpixel and Support Vector Machine（SVM） is proposed.Simple Linear Interactive Cluster（SLIC） is firstly used to divide the image into superpixels.And then,intensity,texture and local features are extracted for each superpixel.At last,SVM is trained based on these superpixels.Because the high-model is trained by introducing the segmentation result information,the high-model can be used for recognizing tumor regions misclassified by low-level segmentation model.In the process of image segmentation by using low-level segmentation model,complex characteristics of the ultrasound images may result in segmentation error of some local tumor regions which appear large intensity difference or texture variation.However,high-level segmentation model has ability to learn certain characteristics of the low-level segmentation model,which can be complementary to low-level segmentation model.Therefore,the high-level segmentation model can be used to correct the segmentation error caused by using low-level segmentation model,which can be used to improve the segmentation performance.We conduct the experiments on our self-constructed database and the experimental results demonstrate the effectiveness and robustness of the proposed method.