摘要
目的:为了提高肺部疾病的临床诊断准确率及其手术成功率,需要对肺气管的影像进行准确的分割。方法:提出一种全新的针对肺气管图像的三维图像分割算法,将深度学习中的对抗生成网络结构(GAN)、密集连接网络模型(Dense Net)以及多尺度连接(Multi Scale)结构应用到临床三维图像的分割中。结果:该方法可以从读取数据块中做到像素级的分割,根据相对坐标位置对分割结果进行投票,结合最大联通分量后处理办法,分割出精确到第四支的气管,算法在测试数据上的分割结果获得交并比(IOU)超过91.4%的准确率。结论:基于对抗生成网络的图像分割算法,无需人工交互即可获得形态学结构复杂(肺气管)的三维目标,能提高肺部疾病的临床诊断准确率和手术成功率。
Objective:It is necessary to segment bronchia image accurately to improve the clinical diagnostic accuracy and surgical success rate of pulmonary diseases.Methods:A new three-dimensional image segmentation algorithm for bronchia images is proposed,and the generative adversarial network structure(GAN),dense connection network model(Dense Net)and multi-scale connection(Multi Scale)structure in deep learning are applied to three-dimensional image segmentation.Results:The method can achieve pixel-level segmentation from the read data block,vote on the segmentation results according to the relative coordinate position,and combine the maximum connectivity component post-processing method to segment the trachea accurate to the fourth branch.The algorithm is testing the data in the above,the segmentation result have more than 91.4%cross-combination accuracy.Conclusion:The image segmentation algorithm based on generative adversarial network can obtain complex morphological structures(bronchia)threedimensional targets without human interaction,which can improve the clinical diagnostic accuracy and surgical success rate of pulmonary diseases.
出处
《中国数字医学》
2021年第10期93-97,共5页
China Digital Medicine
基金
国家自然科学基金(编号:61271336)
厦门市科技计划(编号:3502Z20209154).
关键词
深度学习
对抗生成网络模型
肺气管图像分割
deep learning
generative adversarial network model
bronchia image segmentation
作者简介
通信作者:郭明,陆军第七十三集团军医院心胸外科,361003,福建省厦门市思明区文园路92-96号,E-mail:13358378777@163.com
