| 探讨CT扫描剂量对人工智能检测肺结节效能的影响 |
| |
| 引用本文： | 温德英,潘雪琳,姚辉,李吉杰,邓巧,唐露,伍希,孙家瑜.探讨CT扫描剂量对人工智能检测肺结节效能的影响[J].CT理论与应用研究,2021,30(4):455-465. |
| |
| 作者姓名： | 温德英 潘雪琳 姚辉 李吉杰 邓巧 唐露 伍希 孙家瑜 |
| |
| 作者单位： | 1. 四川大学华西医院放射科, 成都 610041; |
| |
| 基金项目： | 四川省科技厅重点研发项目（2020YFS0123）。 |
| |
| 摘 要： | 目的:探讨影响人工智能检测肺结节效能的因素，力求为不同性质的结节提供个性化的扫描剂量及人工智能系统，同时为各人工智能系统适宜的扫描条件提供参考。方法:标准成人男子胸部X线/CT影像模型，内部随机分布15个不同密度和大小的模拟肺结节，采用不同的管电压和管电流对模型进行扫描，共扫描50次。应用不同公司的人工智能系统进行肺结节检测，采用Pearson χ2检验或Fisher确切概率法比较各组检出率和假阴性率；采用Kruskal-Wallis H检验比较假阳性率。结果:①不同管电压条件下，公司A和公司C对不同性质肺结节的检出率无统计学差异；公司B对+100HU结节的检出率，70kV（100%）组高于120kV（80%）和140kV（80%）组；公司B对3mm结节的检出率，70kV组（33.33%）高于120kV（0%）和140kV（0%）组，差异有统计学意义。②各管电压组内不同管电流间及各管电压组间，检出率、假阴性率的差异无统计学意义。各管电压组间假阳性率的差异具有统计学意义。③公司A在70kV组检出率（64.44%）低于公司B（80.00%）、假阴性率（35.56%）高于公司B（20.00%）；公司A的假阳性率高于公司B和公司C；公司B和公司C间检出率、假阴性率、假阳性率无统计学差异。结论:人工智能辅助肺结节检测的灵敏度与CT扫描剂量无关，与结节性质及AI系统性能有关。本研究中公司B和公司C整体性能高于公司A，最佳扫描管电压分别是70kV、70kV和100kV。
|
| 关 键 词： | 肺结节 人工智能 影像模型 |
| 收稿时间： | 2021-04-29 |
To Explore the Effect of CT Scan Dose on the Efficacy of Artificial Intelligence in Detecting Lung Nodules |
| |
| WEN Deying,PAN Xuelin,YAO Hui,LI Jijie,DENG Qiao,TANG Lu,WU Xi,SUN Jiayu.To Explore the Effect of CT Scan Dose on the Efficacy of Artificial Intelligence in Detecting Lung Nodules[J].Computerized Tomography Theory and Applications,2021,30(4):455-465. |
| |
| Authors: | WEN Deying PAN Xuelin YAO Hui LI Jijie DENG Qiao TANG Lu WU Xi SUN Jiayu |
| |
| Affiliation: | 1. Department of Radiology, West China Hospital, Sichuan University, Chengdu 610041, China;2. Department of Medical Record, West China Second Hospital, Sichuan University, Chengdu 610041, China |
| |
| Abstract: | Objective: To explore the factors that affect the effectiveness of artificial intelligence systems in detecting lung nodules, and strive to provide personalized scanning conditions and artificial intelligence systems for nodules of different natures, and provide references for the appropriate scanning conditions of various artificial intelligence systems. Methods: Standard adult male chest X-ray/CT image model, in which 15 simulated lung nodules with different density and size are randomly distributed was scanned by different tube voltages and tube current at a total of 50 times. The artificial intelligence systems of different companies are used to detect lung nodules. The Pearson χ2 test or Fisher exact probability method is used to compare the detection rate and false negative rate of each group; the Kruskal-Wallis H test is used to compare the false positive rate. Results: (1) Under different kV conditions, the detection rates of nodules of different densities and sizes of company A and C were not statistically different; the detection rate of +100 nodules in company B is higher in the 70kV (100%) group than that in the 120kV (80%) and 140kV (80%) group; the detection rate of 3mm nodules in company B was higher in the 70kV group (33.33%) than in the 120kV (0%) and 140kV (0%) group. (2) There was no significant difference in detection rate and false negative rate among different mAs in each kV group of the three companies or each kV group. The difference in false positive rate among the kV groups was statistically significant. (3) The detection rate of company A in the70kV group (64.44%) is lower than that of company B (80.00%), and the false negative rate (35.56%) is higher than that of company B (20.00%); The false positive rate of company A is higher than that of company B and company C. There is no statistical difference in the detection rate, false negative rate, and false positive rate between company B and C. Conclusions: The sensitivity of artificial intelligence-assisted lung nodule detection has nothing to do with the CT scan dose, but is related to the nature of the nodule and the performance of the AI system. In this study, the overall performance of company B and C is higher than that of company A, and the best scanning tube voltages are 70kV, 70kV, and 100kV respectively. |
| |
| Keywords: | |
| 本文献已被 CNKI 万方数据 等数据库收录！ |
| 点击此处可从《CT理论与应用研究》浏览原始摘要信息 |
|
点击此处可从《CT理论与应用研究》下载全文 |
|
