文章摘要

文章信息/Info

Title:: Robust segmentation of lung parenchyma based on fuzzy region contrast enhancement

作者:: 武振宇1; 白培瑞1△; 刘艺炜1; 任延德2; 1.山东科技大学电子通信与物理学院，青岛 266590;2青岛大学附属医院放射科,青岛 265000

Author(s):: WU Zhenyu1; BAI Peirui1; LIU Yiwei1; REN Yande2; 1.College of Electronic, Communication and Physics, Shandong University of Science and Technology, Qingdao 266590, China； 2.Department of Radiology, Affiliated Hospital of Qingdao University, Qingdao 265000

Keywords:: Lung parenchyma segmentation; Super-pixel; Local contrast enhancement; Refinement segmentation; Morphological processing

摘要:: 基于阈值操作的肺实质分割对CT图像的对比度敏感，常常造成肺部粘连区域的肺实质分割失败。提出一种融合模糊区域对比度增强与阈值和形态学细化分割的新的肺实质分割算法。首先，根据图像的灰度信息利用线性迭代聚类将图像预分割为多个超像素。然后，根据超像素的灰度统计信息自动定位模糊区域，并进行自适应对比度增强。最后，基于阈值和形态学操作进行细化分割，准确提取肺部粘连区域和肺实质。通过对kaggle肺部数据集30位患者的300张CT图像进行测试，结果表明本研究算法的平均分割准确率（Dice系数）为98.65%，过分割率为0.21%，欠分割率为1.33%，整体分割性能比传统阈值操作和形态学方法有明显提升。

Abstract:: In this paper, we proposed a novel lung parenchyma segmentation algorithm which is to combine contrast enhancement of fuzzy region with refinement segmentation using threshold and morphological. This algorithm could deal with effectively negative effects of lung adhesion region to lung parenchyma segmentation. First, the original CT image was pre-segmented into several super-pixel patches using the linear iterative clustering（SLIC0）in terms of gray intensity. Second, the fuzzy regions on CT image were located automatically by statistic information of the super-pixel patches, and contrast enhancement was implemented adaptively in the corresponding regions. Finally, refinement segmentation was performed by employing thresholding and morphological operation to extract the lung adhesion regions and lung parenchyma accurately. The performance of the proposed algorithm was validated on 300 CT images of 30 patients which were obtained from the open lung dataset, i.e. kaggle. The experimental results demonstrate that the mean dice coefficient of the proposed algorithm is 98.65%, the mean over-segmentation is 0.21% and the mean under-segmentation is 1.33% respectively. The segmentation performance of the proposed algorithm outperforms obviously the classical threshold operation and morphological methods.