基于图像处理技术的肺结节分割和肺部良恶性分型

Segmentation of pulmonary nodules and classification of benign and malignant based on image processing techniques

目的探讨CAD技术和图像处理算法应用于肺结节分割和肺部良恶性分型的效果。方法选择2015年1月至2020年12月在我院数据库中肺结节CT图像2856个进行回顾性分析,其中直径标准为3~30 mm。将其中的2744个肺结节作为训练集,良性和恶性分别为2380、364个;剩下的112个作为验证集,良性和恶性分别为62、50个。通过PyRaDiomics提取各个结节的影像学特征,并应用三联法或者LASSO算法进行特征筛选,以构建出基于SVM算法的肺结节良恶性预测模型,评价分析最优模型在验证集中的效能,同时采用交叉验证法评价训练集。结果通过训练样本重复进行特征选择,结果显示组间对比的特征差异为820,具有显著性意义(P<0.05),并应用LASSO确定了特征17个;通过无监督聚类观察验证集,本实验构建的分类模型在肺部良恶性分型中的准确率达到了69.7%,其中特异度0.680、敏感度0.651,NPV、PPV分别为0.680、0.707;在数据库验证集中进行验证,应用LASSO算法得出最优诊断模型的特异度为0.711、符合率为0.756、敏感度为0.811、AUC为0.776、PPV和NPV分别为0.690、0.819。结论CAD技术和图像处理算法在肺结节分割和肺部良恶性分型中应用效果较好,具有一定的泛化性,可以进一步在计算机辅助诊断中推广。

Objective To explore the application of computer aided diagnosis(CAD)technology and image processing algorithm in pulmonary nodule segmentation and pulmonary benign and malignant classification.Methods 2856 CT images of pulmonary nodules in our hospital database from January 2015 to December 2020 were retrospectively analyzed,of which the diameters were 3-30 mm.Taking 2744 pulmonary nodules as the training set,the benign and malignant nodules were 2380 and 364 respectively;The remaining 112 were used as the validation set,and the benign and malignant were 62 and 50 respectively.The imaging features of each nodule are extracted by pyradiomics,and the triad method or LASSO algorithm is used for feature screening,so as to build a lung nodule benign and malignant prediction model based on SVM algorithm,to evaluate and analyze the effectiveness of the optimal model in the verification set,and use the cross verification method to evaluate the training set.Results Through repeated feature selection of training samples,the results showed that the feature difference between the groups was 820,the comparison was significant(P<0.05),and 17 features were determined by LASSO;Through the unsupervised cluster observation verification set,the accuracy in the classification of benign and malignant lung reached 69.7%,in which the specificity was 0.680,the sensitivity was 0.651,and the NPV and PPV were 0.680 and 0.707 respectively;The specificity of the optimal diagnostic model was 0.711,the coincidence rate was 0.756,the sensitivity was 0.811,the AUC was 0.776,and the PPV and NPV were 0.690 and 0.819 respectively.Conclusion The application of CAD technology and image processing algorithm in pulmonary nodule segmentation and pulmonary benign and malignant classification has good effects,and has certain general features,which can be further popularized in computer-aided diagnosis.