The availability of pulmonary nodules in CT scan image of lung does not completely specify cancer. The noise in an image and morphology of nodules, like shape and size has an implicit and complex association with cancer, and thus, a careful analysis should be mandatory on every suspected nodules and the combination of information of every nodule. In this paper, we introduce a “denoising first” two-path convolutional neural network (DFD-Net) to address this complexity. The introduced model is composed of denoising and detection part in an end to end manner. First, a residual learning denoising model (DR-Net) is employed to remove noise during the preprocessing stage. Then, a two-path convolutional neural network which takes the denoised image by DR-Net as an input to detect lung cancer is employed. The two paths focus on the joint integration of local and global features. To this end, each path employs different receptive field size which aids to model local and global dependencies. To further polish our model performance, in different way from the conventional feature concatenation approaches which directly concatenate two sets of features from different CNN layers, we introduce discriminant correlation analysis to concatenate more representative features. Finally, we also propose a retraining technique that allows us to overcome difficulties associated to the image labels imbalance. We found that this type of model easily first reduce noise in an image, balances the receptive field size effect, affords more representative features, and easily adaptable to the inconsistency among nodule shape and size. Our intensive experimental results achieved competitive results.

肺部CT扫描图像中肺结节的可用性尚不能完全确定癌症。结节的图像和形态中的噪声（如形状和大小）与癌症之间存在隐式和复杂的关联，因此，必须对每个可疑结节以及每个结节的信息组合进行仔细的分析。在本文中，我们介绍了一种“降噪优先”两径卷积神经网络（DFD-Net）来解决这种复杂性。引入的模型由端对端的去噪和检测部分组成。首先，在预处理阶段采用残差学习降噪模型（DR-Net）去除噪声。然后，采用两径卷积神经网络，该网络以DR-Net的去噪图像作为输入来检测肺癌。两条路径着重于局部和全局特征的联合集成。为此，每个路径采用不同的接收域大小，这有助于对局部和全局依赖性进行建模。为了进一步改善模型性能，以不同于直接将不同CNN层的两组特征进行级联的常规特征级联方法，我们引入了判别相关分析以级联更多具有代表性的特征。最后，我们还提出了一种再训练技术，该技术可使我们克服与图像标签不平衡相关的困难。我们发现，这种类型的模型可以轻松地首先降低图像中的噪声，平衡接收场大小的影响，提供更具代表性的功能，并易于适应结节形状和大小之间的不一致。