赵正凯^1,,
梁勇¹,
周建收¹,
李建林¹,
李娅¹,
程绍玲^2,,
1. 成都市第三人民医院 放射科, 四川 成都 610031
2. 大连医科大学附属第二医院 放射科, 辽宁 大连 116027

详细信息 作者简介: 赵正凯(1990-), 男, 主治医师。E-mail: zzkdoctor@163.com






通讯作者: 程绍玲, 教授。E-mail: doctor_cheng01@sina.com 中图分类号: R445.3


摘要:目的评估基于深度学习的人工智能（artificial intelligence，AI）软件在胸部CT肺磨玻璃结节（ground glass nodule，GGN）检出及定性诊断中的价值。方法收集2018年6月至2020年6月经手术病理确诊的123例GGN患者，共切除154枚GGN。将123例术前高分辨胸部CT图像导入人工智能软件，分别采用人工智能（AI）和影像医师阅片检出GGN并进行良恶性诊断，对比两种方法检出GGN的敏感度、阳性预测值及假阳性率，采用McNemar检验比较影像医师和人工智能对GGN的检出能力。以病理结果为诊断金标准，计算AI与影像医师在恶性GGN诊断中的敏感度、特异度、阳性预测值及阴性预测值。结果123例胸部高分辨CT共检出289枚GGN；AI与影像医师检出GGN的敏感度分别为94.8%、85.1%，阳性预测值分别为94.2%、100%；AI假阳性率为每例胸部CT 0.14个GGN；影像医师无假阳性GGN。AI与影像医师检出全部GGN的能力差异具有统计学意义（P < 0.001）。154枚GGN经手术切除，AI、影像医师及AI联合影像医师诊断恶性GGN的敏感度分别为96.5%、89.4%、98.2%，特异度分别为19.5%、65.9%、69.1%。结论AI检测GGN及诊断恶性GGN的敏感度均高于影像医师，影像医师对于诊断良性GGN优于AI，故AI联合影像医师诊断可以提高GGN的总体诊断效能。
关键词: 肺结节/
人工智能/
计算机体层成像/
定性诊断


Abstract:ObjectiveTo investigate the value of artificial intelligence (AI) based on deep learning in the detection and diagnosis of pulmonary ground glass nodule (GGN) on chest CT.MethodsA total of 123 patients with pulmonary GGN diagnosed by surgery in our hospital from June 2018 to June 2020 were collected, and a total of 154 GGNs were removed. Preoperative high-resolution chest CT images of the 123 patients were imported into the artificial intelligence recognition system, using artificial intelligence (AI) and radiologists to detect and diagnose pulmonary GGN. The sensitivity, positive predictive value and false positive rate of the two diagnostic methods were compared, and McNemar test was used to compare the ability of radiologists and AI to detect pulmonary GGN. With pathological results as the gold standard for diagnosis, the sensitivity, specificity positive and negative predictive values of AI, radiologists and AI combined radiologists in diagnosis of malignant GGN were calculated.ResultsA total of 289 GGNs were detected in 123 cases on high-resolution chest CT. The sensitivity of AI and radiologists to detect GGN was 94.8% and 85.1%, the positive predictive values were 94.2% and 100%, respectively. The false positive rate of AI was 0.14 GGN per chest CT. The radiologist did not have false positive GGN. There was significant difference between AI and radiologists to detect all GGN (P < 0.001). Among them, 154 GGNs were surgically removed. The sensitivity of AI, radiologist and AI combined radiologist to diagnose malignant pulmonary GGN was 96.5%, 89.4% and 98.2%, and the specificity was 19.5%, 65.9% and 69.1%, respectively.ConclusionThe sensitivity of AI in detecting and diagnosing malignant GGN is higher than that of radiologists. Radiologists are better than AI in diagnosing benign GGN. Therefore, AI combined with radiologists can improve the overall diagnostic efficiency of GGNs.
Keywords:pulmonary nodules/
artificial intelligence/
computer tomography/
qualitative diagnosis

PDF全文下载地址:

https://journal.dmu.edu.cn/data/article/export-pdf?id=dlykdxxb_20210309