摘要目的:评价人工智能肺结节辅助诊断系统在肺结节检出及良、恶性鉴别的效能。方法:回顾性分析2016年5月至2020年7月，于兰州大学第二医院胸外科因肺结节就诊的199例患者的临床资料，将术前胸部CT导入人工智能系统，记录检出肺结节的直径、密度分类、恶性风险值。计算人工智能系统对肺结节的检出率，计算人工智能系统在肺结节良恶性鉴别的敏感性、特异性、阳性似然比、阴性似然比，评价其诊断效能并与人工阅片进行比较，以及在肺结节不同大小及密度情况下对肺结节良、恶性鉴别的敏感性及特异性。结果:手术切除取得病理诊断的肺结节共204个，人工智能系统对肺结节的检出率为100%；人工智能系统对肺结节良、恶性鉴别，敏感性95.83%(95% CI：0.8967～0.9883)，特异性25.00%(95% CI：0.1717～0.3425)，阳性似然比1.27(95% CI：1.14～1.44)，阴性似然比0.17(95% CI：0.06～0.46)；人工阅片对肺结节良、恶性鉴别，敏感性87.36%(95% CI：0.7850～0.9352)，特异性72.17%(95% CI：0.6214～0.8079)，阳性似然比3.14(95% CI：2.26～4.37)，阴性似然比0.18(95% CI：0.10～0.31)。5 mm≤肺结节直径<10 mm时，敏感性100%(95% CI：0.6637～1.0000)，特异性50.00%(95% CI：0.01258～0.98740)；10 mm≤肺结节直径<20 mm时，敏感性94.29%(95% CI：0.8084～0.9930)，特异性29.83%(95% CI：0.1843～0.4340)；20 mm≤肺结节直径≤30 mm时，敏感性96.15%(95% CI：0.8679～0.9953)，特异性18.37%(95% CI：0.0876～0.9953)；亚实性肺结节敏感性100%(95% CI：0.9051～1.0000)，特异性20.00%(95% CI：0.0051～0.7164)；实性肺结节敏感性93.22%(95% CI：0.8354～0.9812)，特异性25.24%(95% CI：0.1720～0.3476)。 结论:人工智能肺结节辅助诊断系统在肺结节的检出方面性能较强，但是在肺结节良恶性的鉴别方面不能满足临床要求，现阶段人工智能系统可作为医师的辅助工具，进行肺结节的检出及辅助肺结节的良、恶性判断。

abstractsObjective:To evaluate the efficacy of artificial intelligence assisted pulmonary nodule diagnosis system in detection pulmonary nodule and predicting the malignant probability of pulmonary nodule.Methods:A retrospectively analyze the clinical data of 199 patients with lung nodules in the Thoracic Surgery Department of Lanzhou University Second Hospital from May 2016 to July 2020. The preoperative chest CT was imported into the artificial intelligence system to record the detected lung nodules, to measure nodal diameter and density classification and malignant probability prediction value of each nodule. The detection rate of pulmonary nodules by artificial intelligence system was calculated, and the sensitivity, specificity, positive likelihood ratio and negative likelihood ratio of artificial intelligence system in the differential diagnosis of benign and malignant pulmonary nodules were calculated and compared with manual film reading. and the sensitivity and specificity in the differential diagnosis of benign and malignant pulmonary nodules under the condition of different size and density of pulmonary nodules.Results:A total of 204 pulmonary nodules were pathologically diagnosed by surgical resection, and the detection rate of pulmonary nodules by artificial intelligence system was 100%. The artificial intelligence system can distinguish benign and malignant pulmonary nodules with a sensitivity of 95.83%(95% CI: 0.8967-0.9883), specificity 25.00%(95% CI: 0.1717-0.3425), and a positive likelihood ratio of 1.27(95% CI: 1.14-1.44), negative likelihood ratio 0.17(95% CI: 0.06-0.46), Manual reading for the differentiation of benign and malignant pulmonary nodules has a sensitivity of 87.36%(95% CI: 0.7850-0.9352), specificity 72.17%(95% CI: 0.6214-0.8079), and a positive likelihood ratio of 3.14(95% CI: 2.26-4.37), the negative likelihood ratio is 0.18(95% CI: 0.10-0.31). 5mm≤diameter of pulmonary nodule<10 mm, sensitivity 100%(95% CI: 0.6637-1.0000), specificity 50.00%(95% CI: 0.01258-0.98740), 10 mm≤diameter of pulmonary nodule <20 mm, sensitivity 94.29%(95% CI: 0.8084-0.9930), specificity 29.83%(95% CI: 0.1843-0.4340), 20 mm≤ diameter of pulmonary nodule ≤30 mm, sensitivity 96.15%(95% CI: 0.8679-0.9953), specificity 18.37%(95% CI: 0.0876-0.9953), sensitivity of subsolid lung nodules: 100%(95% CI: 0.9051-1.0000), specificity 20.00%(95% CI: 0.0051-0.7164), solid lung nodule sensitivity 93.22%(95% CI: 0.8354-0.9812), specificity 25.24%(95% CI: 0.1720-0.3476). Conclusion:The artificial intelligence assistant diagnosis system of pulmonary nodules has a strong performance in the detection of pulmonary nodules, but it can not meet the clinical requirements in the differentiation of benign and malignant pulmonary nodules. At present, the artificial intelligence system can be used as an auxiliary tool for doctors to detect pulmonary nodules and assist in the diagnosis of benign and malignant pulmonary nodules.