基于Curvelet变换的肺结节CT图像良恶性分类研究

目的早期肺癌患者的CT图像表现为结节状（在肺野内直径≤3cm的病灶），需要与结核球等良性病变鉴别开，以提高患者的5年生存率。方法本文基于Curvelet变换提取能量、熵、灰度均值及灰度标准差四种纹理特征值，按7：3比例将样本分为训练集与验证集。使用BP（back propagation）神经网络作为分类器。每一种纹理参数测试集的神经网络仿真值结合病理诊断结果绘制受试者工作特征曲线（receiver operator characteristic cllrve，ROC曲线），根据ROC下面积得到最优的几种纹理参数用于良恶性分类，并将分类结果与病理诊断结果进行比较。结果四种纹理参数构建的BP网络均具有诊断价值，每种纹理参数诊断价值各不相同，其中熵与灰度标准差的诊断价值优于能量与灰度均值，并且通过组合多种纹理参数可以提高诊断准确性。结论使用熵与灰度标准差两种纹理特征值构建BP神经网络能达到最好的分类效果，在一定程度上有利于肺癌的早期诊断。

Classification of Malignant and Benign Pulmonary Nodules in CT Image Based on Curvelet Transformation

Objective To raise the 5-year survival rate, the CT detected pulmonary nodules,which size is defined smaller than 30mm, is needed to be distinguished between benign or malignantones. Methods Curvelet transformation was introduced in this paper and four texture features, including energy, entropy, gray scale mean and gray scale standardized deviation, were calculated. The samples were divided into 2 parts, 70% in test set, and the 30% in validation set. A back propagation（BP） artificial neutral network was used as the classifier. The testing set of each texture feature obtained a ROC （ receiver operator characteristic curve） byusing its simulation result of the BP artificial neutral network and pathological diagnosis. The optimal texture features were chosen to predict the characteristic of small solitary pulmonary nodules in the CT images compared with other texture features, it was more proper to use the entropy and standard deviation as parameters to establish the prediction model. Results The BP artificial neutral network established by parameters entropy and standarddeviation provided the best discrimination of the benign and the malignant small solitary pulmonary nodules Conclusions We can profit from the diagnosis of early stage carcinoma of the lung to some extent with the BP artificial neutral network, which utilizes the entropy and standard deviation as parameters.