Abstract: The characteristics of pulmonary nodules are complex, which make it difficult to detect lung nodules in chest radiographs. A classification method of lung nodules based on deep neural network is proposed. In this method, the difference between the brightness and the gray level of the chest is reduced by the consistency of the grayscale of the chest. Secondly, the characteristics of the lung nodules can be extracted by the different data amplification methods, and the lung nodules are identified by the improved neural network frame. The proposed algorithm effectively avoids the loss of the feature part of the image during the segmentation of the chest image, and overcomes the shortcomings of the lung nodule characterized by the complexity of the chest image. The results show that the average accuracy of the thoracic lung nodules is 84.2%. It has a certain application value in the classification and identification of pulmonary nodules in medical chest radiographs.

Key words: convolution neural network, chest radiographs, pulmonary nodules, image classification

摘要： 针对肺结节特征复杂且不明显，难以精确诊断出胸片中是否含有肺结节的问题，提出将深度神经网络应用于肺结节分类识别之中。首先通过将胸片灰度一致化，减少由于不同设备导致胸片亮度与灰度的差异；其次采用不同的数据扩增方法使得深度卷积神经网络可以充分提取肺结节的特征；最后通过改进的神经网络架构对肺结节进行分类识别。提出的算法有效地避免了在对胸片图像进行分割时造成图像特征部分丢失的现象，同时克服了由于胸片图像的复杂造成的肺结节特征不明显的缺点。最终通过实验研究证明胸片肺结节分类识别的平均准确率达到84.2%，在医学胸片肺结节的分类识别领域上具有一定的应用价值。

关键词: 卷积神经网络, 胸片, 肺结节, 图像分类