不同CT成像参数对基于深度学习的智能辅助软件诊断肺结节良恶性效能的影响

Efficacy of Deep Learning Based Computer Aided Diagnosis System in Diagnosing Benign and Malignant Pulmonary Nodules Measured by Different CT Imaging Parameters

目的探讨不同CT成像参数对基于深度学习的智能辅助软件(DL-CAD)诊断肺结节良恶性效能的影响,并确定最优成像参数。资料与方法收集行胸外科手术治疗的126例肺结节患者共169个肺结节,包括良性35个、恶性134个,进行CT平扫和增强扫描。采用不同的图像算法(bone算法和standard算法)和后置迭代重建算法(ASIR-V)(30%ASIR-V和50%ASIR-V)重建图像,并导入DL-CAD进行分析,同时记录肺结节恶性概率值及良恶性诊断结果。以病理结果作为诊断“金标准”,比较不同CT成像参数下的DL-CAD受试者工作特征(ROC)曲线下面积(AUC)及诊断效能。结果DL-CAD对肺结节良恶性的诊断效能在CT增强模式下优于平扫模式(AUC:P<0.001;整体效能:P=0.064),bone算法的诊断效能优于standard算法(AUC:P=0.045;整体效能:P=0.013)。结论CT增强扫描及重建算法会影响DL-CAD对肺结节良恶性的诊断效能,在CT增强模式下,使用bone算法,并应用30%ASIR-V或50%ASIR-V,DL-CAD对肺结节良恶性有较好的诊断效能。

Purpose To evaluate the efficacy of deep learning based computer aided diagnosis system(DL-CAD)in diagnosing benign and malignant pulmonary nodules via different CT imaging parameters,and to confirm the optimal values.Materials and Methods A total of 126 patients with pulmonary nodules(total number of pulmonary nodules,n=169;benign nodules,n=35;malignant nodules,n=134)confirmed by thoracic surgery were retrospectively recruited and underwent non-contrast and contrast-enhanced CT scans.All images were reconstructed with different reconstruction algorithms(including bone algorithms and standard algorithms)and post-adaptive statistical iterative reconstruction V(ASIR-V)(30%ASIR-V and 50%ASIR-V).All the reconstructed CT images were imported to DL-CAD,calculating the malignant probability and exporting the results of benign and malignant of pulmonary nodules.The pathological results after surgery were set as the gold standard for the diagnosis of benign and malignant pulmonary nodules.Area under the receiver operating characteristic(ROC)curves(AUC)and overall diagnosis efficacy(sensitivity,specificity,positive predictive value,negative predictive value and accuracy)of DLCAD were calculated and compared under different CT imaging parameters.Results The diagnostic efficiency of DL-CAD in contrastenhanced CT was significantly better than that of non-contrast CT(AUC:P<0.001;overall performance:P=0.064),and bone reconstruction algorithm was significantly better than the standard one(AUC:P=0.045;overall performance:P=0.013).Conclusion Contrast-enhancement CT scan and reconstruction algorithms can affect the diagnostic efficacy of DL-CAD for benign and malignant pulmonary nodules,of which contrast-enhancement,bone reconstruction algorithms and 30%ASIR-V or 50%ASIR-V have better diagnostic efficacy.