| 引用本文: | 侯智超,杨杨,李晓琴.基于三维卷积神经网络的肺结节探测与定位方法[J].生物信息学,2022,20(1):28-34. |
| HOU Zhichao,YANG Yang,LI Xiaoqin.Detection and location of pulmonary nodules based on 3D convolutional neural network[J].Chinese Journal of Bioinformatics,2022,20(1):28-34. |
|
| 摘要: |
| 提出一种基于三维卷积神经网络对肺部计算机断层扫描图像(CT)进行肺结节自动探测及定位的方法。基于开源数据集LUNA16开展研究,对数据进行像素归一化、坐标转换等预处理,对正样本使用随机平移、旋转和翻转的方式进行扩充,对负样本进行随机采样。搭建了三维卷积神经网络并在训练过程中调整网络参数,直到得到性能最佳的网络。此外还设计了模型在肺部的三维空间中标记肺结节的方法。经测试,模型的敏感性为93.03%,特异性为97.39%,结果表明所提方法能够较为准确地探测并标记结节。 |
| 关键词: 深度学习 三维卷积神经网络 肺结节探测 |
| DOI：10.12113/202012007 |
| 分类号:TP183 |
| 文献标识码:A |
| 基金项目:国家自然科学基金项目(No.61931013, No.81701644, No.11832003)；国家重点研发项目(No.2017YFC0111104). |
|
| Detection and location of pulmonary nodules based on 3D convolutional neural network |
|
HOU Zhichao, YANG Yang, LI Xiaoqin
|
|
(Faculty of Environment and Life, Beijing University of Technology, Beijing 100124, China)
|
| Abstract: |
| A method for automatic detection and localization of pulmonary nodules based on three-dimensional (3D) convolutional neural networks for computed tomography(CT) images of lungs was proposed. Based on the study conducted on the open source dataset LUNA16, the data were pre-processed with pixel normalization and coordinate conversion. Positive samples were expanded using random translation, rotation, and flip, and random sampling was conducted for negative samples. A 3D convolutional neural network was constructed and the network parameters were adjusted during the training process until the best performance was obtained. The model was also designed to label lung nodules in the 3D space of the lung. The sensitivity of the model was tested to be 93.03% and the specificity was 97.39%, indicating that the proposed method can detect and label nodules more accurately. |
| Key words: Deep learning 3D convolutional neural network Pulmonary nodule detection |
