基于深度学习的纵隔淋巴结自动分割研究  

Automatic segmentation of mediastinal lymph nodes based on in-depth learning

作　　者：宋宁宁[1] 陈佳豪 周鑫[1] 王雨荷 田书畅 方莹[2] SONG Ning-ning;CHEN Jia-hao;ZHOU Xin(Department of Clinical Medical Engineering,Nanjing First Hospital,Nanjing Medical University(Nanjing First Hospital),Nanjing 210006,China)

机构地区：[1]南京医科大学附属南京医院(南京市第一医院)临床医学工程处,江苏南京210006 [2]南京医学会学术会务部,江苏南京210006

出　　处：《中国医学装备》2022年第10期1-4,共4页China Medical Equipment

摘　　要：目的:基于深度学习方法使用U-Net网络训练模型,实现纵隔淋巴结自动分割。方法:收集医院呼吸内科提供的294例患者的369幅纵隔淋巴结超声弹性图像,将目标淋巴结图像从原始图像中分割后进行剪裁等预处理后,统一尺寸,输入卷积神经网络中的U-Net网络,使用U-Net网络训练的模型,对纵隔淋巴结进行自动分割。结果:U-Net网络训练模型的深度学习方法,实现了纵隔淋巴结自动分割良好的效果,Dice系数可达到0.968 4,接近于1,纵隔淋巴结自动分割效果良好。结论:使用深度学习方法对纵隔淋巴结进行自动分割,可得到很好的分割效果,对于后续淋巴结的性质判定以及肺癌分期的确定具有重要意义。Objective: To realize automatic segmentation of mediastinal lymph nodes through adopted in-depth learning, which used U-Net network training model. Methods: The ultrasound elastic images of 369 mediastinal lymph nodes from 294 patients that provided by the Respiratory Medicine Department of Hospital were collected.The sizes of the images were unified after the images of aimed lymph glands were segmented from original images and they were cut and were conducted other preprocessing. And then, these images were inputted into the U-Net network of convolutional neural network. U-Net network train model was used to conduct automatically segmented for mediastinal lymph nodes. Results: The in-depth learning method of U-Net network training model could realize favorable effect of automatic segmentation for mediastinal lymph nodes. And the Dice coefficient could reach 0.968 4,which closed to 1, and the effect of automatic segmentation for mediastinal lymph nodes was favorable. Conclusion:The use of the in-depth learning method in automatically segmenting the mediastinal lymph nodes can obtain excellent segmentation effect, which is of great significance for the subsequent determination of the nature of the lymph nodes and the determination of staging lung cancer.

关 键 词：深度学习 卷积神经网络 U-Net网络 分割 纵隔淋巴结 

分 类 号：R446[医药卫生—诊断学]