结合深度卷积神经网络与影像学特征的肺结节良恶性鉴别方法

Method for Identifying Benign and Malignant Pulmonary Nodules Combing Deep Convolutional Neural Network and Hand-Crafted Features

提出一种将卷积神经网络(CNN)学习特征与传统影像学特征结合的肺结节良恶性鉴别方法。首先,从电子计算机断层扫描(CT)图像中分割出肺结节区域,并使用传统机器学习方法提取结节区域的影像学特征;然后,使用截取的肺结节训练3D-Inception-ResNet模型,提取网络学习的CNN特征,组合两类特征,并利用随机森林(RF)模型进行特征选择;最后,采用支持向量机(SVM)、RF等传统分类器对肺结节进行良恶性鉴别诊断。使用LIDC-IDRI数据库中的1036个肺结节进行实验验证,最终所提方法的分类准确率、敏感度、特异度及接受者操作特性曲线(ROC)下面积(AUC)分别达94.98%、90.02%、97.03%及97.43%。实验结果表明,所提方法能准确地判别肺结节的良恶性,并优于大部分主流方法。

Here, we present a method for identifying benign and malignant pulmonary nodules that combines convolutional neural network(CNN)learning features and conventional hand-crafted features. First, the pulmonary nodules area is segmented from computed tomography(CT) images, and traditional machine learning methods are used to extract the image features of the nodule area. Then, the CNN features of network learning are extracted, using the intercepted pulmonary nodules to train the 3 D-Inception-ResNet model, and the 2 kinds of features are combined, the random forest(RF) model is used for feature selection. Finally, support vector machine(SVM) and RF classifier are used to identify benign and malignant pulmonary nodules. The 1036 pulmonary nodules in the LIDC-IDRI database are used for experimental verification. Classification accuracy, sensitivity, specificity, and area under the receiver operating characteristic(ROC) curve(AUC) of the proposed method can reach 94.98%, 90.02%, 97.03%, and 97.43%, respectively. The proposed method can accurately distinguish benign and malignant lung nodules, more effectively than most existing mainstream methods, as shown by the experimental results.