| 人工智能辅助规培医师对不同大小肺实性结节检出效能的初步研究 |
| |
| 引用本文： | 孙婷婷,汪琼,谢梅,范鸿禹,张清,伍建林.人工智能辅助规培医师对不同大小肺实性结节检出效能的初步研究[J].CT理论与应用研究,2020,29(4):465-472. |
| |
| 作者姓名： | 孙婷婷 汪琼 谢梅 范鸿禹 张清 伍建林 |
| |
| 作者单位： | 大连大学附属中山医院放射科, 辽宁 大连 116001 |
| |
| 基金项目： | 大连市领军人才科技项目基金（2015E12SF120） |
| |
| 摘 要： | 目的:探讨人工智能（AI）辅助检测软件对低年资规培医生提高肺实性结节检出效能的临床应用价值。方法:收集200例经CT证实有肺实性结节的CT影像，由2名8年影像诊断工作的主治医师结合AI （SCHOLAR，infervision）共同阅片确定肺实性结节数量，分歧时由第3名从事影像诊断15年以上主任医师会诊，最终确定“金标准”。先由低年资规培医师对上述CT图像独立进行肺结节检测（方法A），2周洗脱期后在AI软件辅助下再进行上述CT图像肺结节检测（方法B）。将方法A和方法B标注结果分别与“金标准”比较，记录真阳结节数、假阳结节数，采用SPSS 20.0数据统计软件比较两组间检测灵敏度、假阳性率差异，P<0.05为差异有统计学意义。结果:与方法A相比，方法B灵敏度明显增加，总的实性结节灵敏度提高65%，而假阳性率降低25%，差异有统计学意义（P<0.05）；应用AI辅助软件对检测4组不同大小（D≤4 mm、4 mm 8 mm）肺实性结节的灵敏度均有提高，分别为78%、38%、27%和13.8%。方法A的FROC曲线上面积（AAC）为0.176，方法B为0.085 2，差异有统计学意义（P<0.05）；方法A和方法B平均每例CT阅片时间分别为411.9 s和319.7 s。结论:AI辅助检测软件可明显提高低年资规培医师对CT上不同大小肺实性结节的检出效能，尤其对直径≤4 mm肺结节的检出更具优势。
|
| 关 键 词： | CT检查 肺结节大小 人工智能 检出效能 |
| 收稿时间： | 2019-12-11 |
Detection Efficiency of Residents Assisted by Artificial Intelligence for Pulmonary Solid Nodules with Different Sizes: A Preliminary Study |
| |
| Affiliation: | Department of Radiology, Affiliated Zhongshan Hospital of Dalian University, Dalian 116001, China |
| |
| Abstract: | Objective:To investigate the clinical application value of artificial intelligence(AI) assisted detection software for improving the detection efficiency of pulmonary solid nodules in inexperienced residents. Methods:A total of 200 CT images of pulmonary solid nodules confirmed by CT were collected. One senior radiologist with more than 8 years' experience read CT images based on the initial diagnosis of another senior radiologist with similar experience and a final decision was subsequently conducted by deputy chief radiologist with more than 15 years' experience to determine the ground truth solid lung nodules. One resident read the images without AI software(method A) and the same resident read CT images with AI software(method B) after two weeks' washout period(method B). The results of methods A and B were compared with the gold standard nodules. The number of true positive nodules and the number of false positive nodules were recorded. The difference between detection sensitivity and false positive rate between the two groups was analyzed by SPSS 20.0. The difference was statistically significant(P < 0.05). Results:Compared with method A, the sensitivity of method B increased significantly, the sensitivity of total solid nodules increased by 65%, and the false positive rate decreased by 25%, the difference was statistically significant(P < 0.05); The sensitivity of the four groups of different sizes(D ≤ 4 mm, 4 mm < D ≤ 6 mm, 6 mm < D ≤ 8 mm, D > 8 mm) for lung solid nodules was improved by AI assisted software, and the increase rate was 78%,38%, 27% and 13.8%, respectively. The area on the FROC curve of method A(AAC) was 0.176, the method B was 0.0852, and the difference was statistically significant(P < 0.05). The average reading time of the two methods A and B was 411.9 seconds and 319.7 seconds respectively. Conclusion:AI assisted software can significantly improve the detection efficiency of inexperienced residents for different sizes of lung solid nodules in CT, especially for the detection of lung solid nodules ≤ 4 mm in diameter. |
| |
| Keywords: | |
| 本文献已被 CNKI 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 |
|
