Abstract: This paper discusses the effects of pulmonary nodule images of different scales and different modes on deep convolutional neural network model. A novel 2D Multi-View Fusion（2D MVF） method is proposed for lung CT image processing which can catch more lung nodule information than the conventional 2D manner and introduce less interfering tissues than the 3D manner. To validate the model, the LIDC-IDRI and the LUNA16 datasets are proposed and four types of lung nodule images are gotten, including 2D, 3D, 2D FVF（Full-view Fusion）, and 2D MVF lung nodule images, and each type of the image contains 16, 25, 36 three different scales. Then, four types of models of 2D CNN, 3D CNN, 2D FVF-CNN, 2D MVF-CNN are constructed. The above data are used to train and verify the model. The experimental results show that 2D MVF mode has better classification performance than other lung nodule image mode，and the classification performance at small scale is relatively better comparing various nodule image scales.

Key words: computer-aided diagnosis, lung nodule classification, convolutional neural network；multi-view fusion；multi-scale and multi-mode image

摘要： 基于深度卷积神经网络模型，讨论了不同尺度及不同模式肺结节图像对模型分类表现的影响，并提出了一种2D多视图融合的肺图像处理方法，该方法比传统的2D方式能获取更多的肺结节信息，同时又能比3D的方式引入更少的干扰组织。为了验证模型，对LIDC-IDRI和LUNA16数据集进行了预处理，得到了16、25、36三种尺度下2D、3D、2D全视图融合以及2D多视图融合四种不同模式的肺结节图像，然后构建了2D CNN、3D CNN、2D全视图融合卷积神经网络、2D多视图融合卷积神经网络四种模型。利用上述样本对模型进行训练和验证，最终结果表明，2D多视图融合模式下的肺结节图像相对于其他模式图像具有更佳的肺结节分类表现；对比多种尺度图像，小尺度下的分类表现相对更佳。

关键词: 计算机辅助诊断, 肺结节分类, 卷积神经网络, 多视图融合, 多尺度多模式图像