摘要
目的:提出一种基于全卷积神经网络的肺结节测量校准方法,以改善肺结节测量过程中由于不同的呼吸屏气造成的测量误差。方法:选择美国国立卫生研究院的CT图像数据集中的263幅胸部图像制作符合全卷积神经网络训练与测试要求的数据集,然后构建用于肺部轮廓勾画的全卷积语义分割网络,再通过对网络的训练实现对肺部轮廓的准确勾画,最后将该网络部署于校准软件中以实现对不同屏气相的肺结节大小的校准。选取25例行PET/CT肿瘤筛查的患者数据,对屏气和自由呼吸2种模式下的胸部扫描图像进行肺结节测量以验证校准效果。结果:全卷积神经网络对肺轮廓分割的分割效果较好,交并比为0.8362,准确率为0.9247,平均边界得分为0.7169。25例患者中有16例2种呼吸模式的测量结果差异较大,其中9例(占比56.25%)通过该校准方法缩小了测量结果差异。结论:该校准方法实现了对肺结节物理尺寸测量结果的校准,能提高对肺结节诊断的准确性和效率。
Objective To propose a fully convolutional neural network-based calibration method for lung nodule measurement to eliminate the errors caused by breath holding during lung nodule measurement.Methods The 263 chest images from the National Institutes of Health CT image dataset were selected to create a dataset that met the training and testing requirements for a fully convolutional neural network,and then a fully convolutional semantic segmentation network was constructed for lung contouring,trained to accurately outline the lung contours and finally deployed in the calibration software to calibrate the lung nodule size for different breath-holding phases.Data from 25 patients undergoing PET/CT oncology screening were selected and lung nodules were measured in two modes of breath-holding and free-breathing to verify the calibration effect.Results The fully convolutional neural network segmented the lung contours well,with an intersection ratio of 0.8362,an accuracy of 0.9247 and a mean boundary score of 0.7169.Sixteen of the 25 patients had large differences between the measurements of the 2 respiratory modes,9 of which(56.25%)had their measurement differences reduced by this calibration method.Conclusion The calibration method designed enables the calibration of the physical dimension measurement of lung nodules and thus improves the accuracy and efficiency of their diagnosis.
作者
杨静
汤福南
张晖
YANG Jing;TANG Fu-nan;ZHANG Hui(Department of Radiology,Children's Hospital Affiliated to Nanjing Medical University,Nanjing 210008,China;Department of Medical Engineering,the First Affiliated Hospital with Nanjing Medical University,Nanjing 210029,China)
出处
《医疗卫生装备》
CAS
2022年第1期17-21,31,共6页
Chinese Medical Equipment Journal
关键词
肺癌
肺结节
肺结节测量
全卷积神经网络
语义分割
lung cancer
lung nodule
lung nodule measurement
fully convolutional network
semantic segmentation
