计算机科学与探索 ›› 2020, Vol. 14 ›› Issue (8): 1358-1367.DOI: 10.3778/j.issn.1673-9418.2001042

• 图形图像 • 上一篇    下一篇

改进的卷积神经网络在肺部图像上的分割应用

钱宝鑫,肖志勇,宋威   

  1. 江南大学 人工智能与计算机学院,江苏 无锡 214122
  • 出版日期:2020-08-01 发布日期:2020-08-07

Application of Improved Convolutional Neural Network in Lung Image Segmentation

QIAN Baoxin, XIAO Zhiyong, SONG Wei   

  1. School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2020-08-01 Published:2020-08-07

摘要:

CT成像技术是辅助医生诊断肺部疾病的重要手段。针对肺部各组织结构复杂,难以准确地对肺部CT图像中肺实质进行分割和提取的问题,提出了一种编/解码模式的肺分割算法。为了获得图像的多尺度信息,首先向网络模型中输入多尺度图像,使用残差网络结构作为编码模块,在扩展网络深度的同时不造成网络退化问题;此外,在编码和解码之间利用空洞空间金字塔池化(ASPP)充分提取上文多尺度信息;最后利用级联操作,将捕捉到的信息与编码层信息级联,结合注意力机制从而提高分割精度。通过对LUNA16数据集中89位患者的13 465张CT图像进行测试,以相似性系数和精确度作为主要评判标准,实验精度分别达到了99.56%和99.33%。实验结果表明,该方法能有效分割出肺实质区域,与其他网络相比分割效果更好。

关键词: CT图像, 肺实质, 医学图像分割, 深度学习

Abstract:

CT imaging technology is an important means to assist doctors in diagnosing lung diseases. In order to solve the problem that the lung parenchyma is difficult to be segmented and extracted accurately from the CT images of the lung due to the complex structure of the lung tissues, an algorithm of lung segmentation with encoding/decoding mode is proposed. In order to obtain the multi-scale information of the image, the multi-scale image is firstly input into the network model, and the residual network structure is used as the coding module to expand the depth of the network without causing network degradation. In addition, the multi-scale information is fully extracted by using atrous spatial pyramid pooling (ASPP) between encoding and decoding. Finally, the cascade operation is used to cascade the captured information with the information of the coding layer, and the attention mechanism is combined to improve the segmentation accuracy. By testing 13465 CT images of 89 patients in the LUNA16 data set, the experimental accuracy reaches 99.56% and 99.33%, respectively, with the similarity coefficient and accuracy as the main evaluation criteria. Experimental results show that the method can effectively segment the lung parenchyma area, and the segmentation effect is better than other networks.

Key words: CT image, lung parenchyma, medical image segmentation, deep learning