Abstract:

The early detection and diagnosis of lung cancer is the key to improve the survival rate of lung cancer patients. Due to the small nodules in the early stage of lung cancer, the existing pulmonary nodules detection system is easy to miss the diagnosis when detecting these nodules. Accurate detection of early lung nodules is crucial to improve the cure rate of lung cancer. In order to reduce the missed rate of early nodules in the detection system, it is necessary to optimize the extraction procedure of candidate nodules. The shortcut of residual network is introduced into U-Net, which effectively solves the disadvantage of poor results caused by the lack of depth in the traditional U-Net. On the basis of this improvement, a U-type noise residual network NRU（Noisy Residual U-Net） is proposed to enhance the sensitivity of neural network to small nodules by using the characteristics of hopping layer connection and adding noise to the convolutional layer. The Lung Nodule Analysis 2016 and Alibaba Tianchi lung cancer detection competition data sets are used to train the neural network. Comparison experiments between U-Net and NRU show that the sensitivity of the algorithm to small nodules with a diameter of 3-5 mm （97.1%） is greater than that of U-Net（90.5%）.

Key words: lung cancer, pulmonary nodules, pulmonary nodules detection system, noise, residual network

摘要：

肺癌的早期发现和早期诊断是提高肺癌患者生存率的关键。由于肺癌早期结节很小，目前已有的肺结节检测系统在检测这些结节时很容易漏诊。准确检测早期肺癌结节对于提高肺癌治愈率至关重要，为了降低检测系统对早期结节的漏诊率，需要优化候选结节的提取步骤。在U-Net网络中引入残差网络的捷径，有效解决了传统U-Net网络由于缺乏深度而导致结果较差的问题。在此改进的基础上提出了一种U型噪声残差网络NRU（Noisy Residual U-Net），通过利用跳跃层连接的特性和向卷积层添加噪声来增强神经网络对小结节的灵敏度。使用Lung Nodule Analysis 2016和阿里巴巴天池肺癌检测竞赛数据集训练神经网络。U-Net和NRU之间的比较实验表明，该算法对直径为3~5 mm（97.1％）的小结节的灵敏度大于U-Net值（90.5％）。

关键词: 肺癌, 肺结节, 肺结节检测系统, 噪声, 残差网络