胸部增强CT不同重组算法深度学习诊断肺结节良恶性  

Malignancy of pulmonary nodules at contrast-enhanced CT:effect of different reconstruction kernels on deep learning based computer-aided diagnosis system performance

作　　者：沈晶[1] 林琳[1] 刘文飞[1] 伍建林[1] SHEN Jing;LIN Lin;LIU Wen-fei;WU Jian-lin(Department of Radiology,Affiliated Zhongshan Hospital of Dalian University,Liaoning 116001,China)

机构地区：[1]大连大学附属中山医院放射科,辽宁大连116001

出　　处：《影像诊断与介入放射学》2021年第1期25-28,共4页Diagnostic Imaging & Interventional Radiology

摘　　要：目的探讨不同重组算法CT增强深度学习对肺结节良恶性诊断的差异性。方法收集从2018年1月~2019年1月40例肺结节患者的CT增强扫描资料。用不同的重组算法(B30f和B70f)重组1 mm薄层CT图像。所有病例均经手术病理证实,其中良性结节10例,恶性30例。每个结节用深度学习智能辅助诊断系统(InferRead Lung CT Research,Infervision,Beijing,China)分析其恶性概率值(0~100%),在B30f和B70f重组算法下,记录每个结节的恶性概率值,并与病理结果进行比较。计算不同重组算法良恶性结节诊断符合率的ROC曲线下面积(AUC)。结果B30f重组对良恶性结节的诊断符合率分别为42.86%和90.00%;B70f重组对良恶性结节的诊断符合率分别为44.44%和86.67%。同时,与B30f(AUC为0.473)相比,B70f预测所有病变的AUC为0.704。结论应用DL-CAD系统,B70f重组对良恶性肺肿瘤的检出率明显提高。利用基于不同的重组算法增强CT图像DL诊断系统有助于发现和预测不同恶性风险的肺结节。Objective The purpose of this study was to evaluate the effect of different reconstruction kernels on deep learning based computer-aided diagnosis(DL-CAD)system detection of benign and malignancy pulmonary nodules at contrast-enhanced CT.Methods Contrast-enhanced CT images were collected from 40 patients with pathologically confirmed with benign(10)and malignant(30)lung nodule between January 2018 and January 2019.Image sets were simulated with different reconstruction kernel(B30f and B70f)at 1-mm thickness.Each nodule was screened using a DL-CAD system(InferRead Lung CT Research,Infervision,Beijing,China)and assigned with 0-100%probability of malignancy.Using B30 and B70 reconstruction algorithms,we recorded the probability value of malignancy per nodule and compared with pathological diagnosis.Accuracy in categorizing benign and malignant nodules and the area under the receiver operating characteristic curve(AUC)for different reconstruction kernels were also calculated.Results For reconstruction kernel of B30,diagnostic accuracy was 42.86%for benign and 90.00%for malignant nodules.B70 demonstrated accuracy of 44.44%for benign and 86.67%for malignant nodules.B70 had greater AUC of 0.704 than B30(0.473)for predicting all lesions.Conclusion Reconstruction kernel of B70 with the use of DL-CAD system is superior to B30 for differentiating benign from malignant lung nodules.

关 键 词：深度学习 重组算法 肺结节 

分 类 号：R734.2[医药卫生—肿瘤] R730.44[医药卫生—临床医学]