【作者】 黄旭东；

【导师】 姜杰；

【作者基本信息】 厦门大学 ， 外科学， 2022， 硕士

【摘要】 目的:探讨人工智能(AI)算法在与低剂量计算机断层扫描(LDCT)下肺结节筛查的性能水平,它们如何与高年资医生进行比较,以及AI算法和低年资医生如何相互补充。方法:回顾性研究分析2020年9月至2021年9月厦门大学附属第一医院肺结节MDT收治入胸外科行肺手术的患者为研究对象。然后分别用AI肺结节检测系统(AI组)、高年资医生组和低年资医生组进行分析,对每组肺结节的检出率、漏检率、误检率以及对比术后病理回报的诊断准确率进行分析,以高年资医生组诊断为标准,分别计算AI组及低年资医生组对肺结节诊断的敏感性、特异性,评估AI的检测效能。结果:低年资医师组在对结节数目判断上:漏诊45个,漏诊率13.9%,误诊数42个,误诊率13.0%。AI组在对结节数目判断上:漏诊21个,漏诊率6.5%,误诊数32个,误诊率9.9%.低年资医师借助AI算法的帮助下较独立阅片可以多检测出33个(73.3%)漏诊的结节,同时也可以纠正其中27个(64.3%)误诊的结节,检测效能整体有所提高且具有统计学意义(P<0.05)。在对结节良恶性分析判断上:良性结节组与恶性结节组间的“性别”、“年龄”、“结节位置”、“最大直径”方面没有统计学意义(P>0.05),但在“实性成分占比”(18.25%±14.25%VS30.58%±18.58%,P<0.01),“恶性预测概率”(46.69%±19.22%VS 73.25%±11.42%,P<0.01)方面具有统计学意义。应用ROC曲线分析肺结节“最大直径”、“实性成分占比”、“恶性预测概率”的数据如下:“恶性概率预测”ROC曲线下面积AUC=0.880,截断值为58.5%;“实性成分占比”ROC曲线下面积AUC=0.723,截断值为20.5%;“最大直径”ROC曲线下面积AUC=0.574,截断值为9.5mm;将AI对结节预测结果与高年资医师相比较,其结果一致性较高(KAPPA=0.713)。在评估肺腺癌恶性程度时,使用“实性成分占比”和“恶性预测概率”两项指标预测AIS的ROC曲线:“实性成分占比”曲线下面积AUC=0.862,截断值为21.5%,“恶性概率预测”曲线下面积AUC=0.674,截断值为72.5%,表明该两项指标可较为准确预测AIS。使用“实性成分占比”指标预测MIA的R0C曲线下面积AUC=0.661,截断值为36%,表明这项指标不能准确预测MIA。使用“实性成分占比”指标预测IAC的ROC曲线下面积AUC=0.941,截断值为36%,表明这项指标可较为准确预测IAC。结论:目前AI算法在评估肺结节方面已可接近两名高年资医师水平,尽管目前人工智能算法(AI)对肺结节检测仍有一定的检测误差,但整体仍能提高低年资医生的检测水平,提升检测效率,并且对早期肺腺癌亚型的鉴别具有重要价值。更多还原

【Abstract】 Objective:To investigate the performance of artificial intelligence(AI)algorithm in pulmonary nodule screening under low-dose computed tomography(LDCT).How they are compared with senior doctors,and how AI algorithm and junior doctors complete each other.Methods:A retrospective study was conducted to analyze the patients who were admitted to the Department of Thoracic Surgery with MDT for pulmonary nodules in the First Affiliated Hospital of XiaMen University and underwent pulmonary surgery from September 2020 to September 2021 as the research objects.Then,the AI pulmonary nodule detection system(AI group),the senior doctor group and the junior doctor group were used to analyze the detection rate,missed detection rate,false detection rate of each group of pulmonary nodules and the comparison of postoperative The diagnostic accuracy rate of pathological returns was analyzed.Taking the diagnosis of the senior doctor group as the standard,the sensitivity and specificity of the AI group and the junior doctor group for the diagnosis of pulmonary nodules were calculated respectively,and the detection efficiency of the AI was evaluated.Results:In the judgment of the number of nodules in the junior physician group,45 were missed,with a missed diagnosis rate of 13.9%,and 42 were misdiagnosed,with a misdiagnosis rate of 13.0%.The AI group judged the number of nodules:21 were missed,the missed diagnosis rate was 6.5%,the number of misdiagnoses was 32,and the misdiagnosis rate was 9.9%.With the help of the AI algorithm,junior doctors can detect 33 more nodules than independent image reading(73.3%)of the missed nodules,and 27(64.3%)of them could be corrected,and the overall detection performance was improved with statistical significance(P<0.05).In the analysis and judgment of benign and malignant nodules:there was no statistical significance in "sex","age","nodule position"and "maximum diameter" between the benign nodule group and the malignant nodule group(P>0.05).However,there is some statistical significance in terms of "proportion of solid components"(18.25%±14.25%VS30.58%±18.58%,P<0.01)and "probability of malignant prediction"(46.69%±19.22%vs 73.25%±11.42%,P<0.01).The data of "maximum diameter","solid component proportion",and "malignancy prediction probability" of pulmonary nodules were analyzed by ROC curve as follows:the area under the ROC curve of "malignant probability prediction" AUC=0.880,and the cut-off value was 58.5%;The area under the ROC curve of "proportion of sex components" AUC=0.723,the cut-off value is 20.5%;the area under the ROC curve of "maximum diameter" AUC=0.574,the cut-off value is 9.5mm;the AI prediction results of nodules were compared with postoperative pathology,the results were highly consistent(KAPPA=0.713).When evaluating the degree of malignancy of lung adenocarcinoma,the ROC curve of AIS was predicted using two indicators:"solid component proportion"and "malignancy prediction probability":the area under the "solid component proportion" curve AUC=0.862,and the cutoff value was 21.5%,the area under the curve of "prediction of malignant probability" was AUC=0.674,and the cut-off value was 72.5%,indicating that these two indicators can accurately predict AIS.The area under the ROC curve of predicting MIA using the "solid component proportion" index was AUC=0.661,and the cut-off value was 36%,indicating that this index could not accurately predict MIA.The area under the ROC curve of predicting IAC using the "solid component proportion" indicator was AUC=0.941,and the cut-off value was 36%,indicating that this indicator can more accurately predict IAC.Conclusion:Although the artificial intelligence algorithm(AI)still has some errors in the detection of pulmonary nodules,it can improve the detection level of doctors with low experience and improve the detection efficiency.更多还原