一种基于轮廓监督的动态规划乳腺X线影像肿块分割算法

A Dynamic Programming Algorithm for Mass Segmentation on Mammograms Based on Contour Supervision

肿块分割是基于乳腺X线影像的计算机辅助诊断系统的重要环节。良好的分割结果能够更好地反映肿块的病理特征,为后续可疑区域的特征提取和分类提供依据。已有大量文献探讨肿块的分割算法。基于动态规划(DP)的肿块分割算法结合了肿块的边缘信息,以及灰度和大小等先验知识。传统的基于DP的算法自适应性和鲁棒性不足。为克服这些缺点,提出一种基于轮廓监督的动态规划方法,该方法首先使用对比度受限的自适应直方图均衡增强肿块感兴趣区(ROI)的对比度,并使用高斯掩膜掩蔽外围组织;然后,将肿块ROI变换到极坐标,结合肿块的边缘、灰度和大小信息计算局部代价,并根据局部代价矩阵计算累积代价矩阵;最后,在基于动态规划的轮廓跟踪过程中,引入轮廓监督机制,避免周围组织和对比度不足的影响。本文对比了改进后算法与传统算法的分割效果。实验结果表明,高斯掩蔽和轮廓监督的引入,有效地掩蔽了肿块周围组织,避免了轮廓偏离。该算法提高了肿块分割准确性,且具有更好的鲁棒性。

Mass segmentation plays a crucial role in Computer- Aided Diagnosis （CAD） systems on mammograms. Good segmentation result can better reflect the pathological characteristics of mass. And it can provide basis for the subsequent feature extraction and classification of suspicious region. At present, lots of literatures of mass segmentation have been proposed. Dynamic Programming （DP） based mass segmentation algorithm uses edge information as well as the priori knowledge of grey level and size information of mass. But the traditional DP based algorithm has low adaptability and robustness in mammograms. In order to overcome these shortcomings, an improved dynamic programming method was presented. Firstly, the Contrast Limited Adaptive Histogram Equalization （CLAHE） was used to improve contrast of Region of Interest （ROI） about mass. And the Gaussian mask was used to mask the surrounding tissues. Then, transform the ROI into polar coordinate and calculate the local cost matrix combining the edge, grey level and size information of mass. And the cumulative cost matrix was calculated using the local cost matrix. Finally, in the process of contour tracing based on dynamic programming, the contour supervision mechanism was introduced to avoid the contour departure, which is caused by the low contrast and the influence of surrounding tissues, The segmentation results of improved algorithm are compared with the traditional algorithm. The experimental results show that the introduction of Gaussian mask and contour supervision can effectively mask the surrounding tissue and avoid the contour departure. The proposed algorithm improved the segmentation accuracy. It is more robust than the conventional methods.