基于深度学习结合解剖学注意力机制的肺结节良恶性分类

肺结节作为肺癌的初期表现，及时的发现和准确的良恶性诊断对于疾病的治疗具有重要的意义。为了提高肺部CT图像中肺结节良恶性的诊断率，提出一种基于3D ResNet的卷积神经网络，并通过加入解剖学注意力模块有效地提高了肺结节良恶性的分类精度。此外，该方法通过自动分割以获取注意力机制所需的感兴趣区域，实现整个流程的全自动化。解剖学注意力的添加能更好地捕捉图像中的局部纹理信息，进一步提取对于肺结节良恶性诊断有用的特征。本文方法在LIDC-IDRI数据集上进行验证。实验结果表明与传统的3D ResNet及其他现有的方法相比，本文方法在分类精度上有显著的提高，在独立测试集上的最终分类的AUC达到0.973，准确率为0.940。由此可见，本文方法能在辅助医生对肺结节的诊断中起到重要作用。

Classification of benign and malignant pulmonary nodules by deep learning with anatomy-based attention mechanism

Abstract: Pulmonary nodule is the initial sign of lung cancer, and the timely detection and accurate diagnosis of malignant and benign nodules have great significance for the treatment of diseases. In order to improve the diagnostic accuracy of benign and malignant pulmonary nodules in pulmonary CT images, a novel convolution neural network based on 3D ResNet is proposed, with anatomy-based attention mechanism for greatly improving the classification accuracy of pulmonary nodules. In addition, the region of interest required for the attention mechanism is obtained by automatic segmentation, thereby achieving the full automation of the whole process. The addition of anatomy-based attention mechanism can better capture local texture information in CT images and further extract useful features for diagnosing benign and malignant pulmonary nodules. The proposed method is verified in LIDC-IDRI data set. The results show that the proposed method greatly improves classification accuracy as compared with other existed methods and traditional 3D ResNet, achieving an area under the receiver operating characteristic curve (AUC) of 0.973 and an accuracy of 94.0% in the independent test set. The proposed method has the potential to assist doctors in the diagnosis of pulmonary nodules.