基于CT残差神经网络-101-金字塔网络模型鉴别肺良、恶性结节  

Residual neural network-101-feature pyramid network model based on CT for differentiating benign and malignant lung nodules

作　　者：刘刚[1] 解晓婷 何慧[1] 刘飞 毛旭[1] 桑菁遥 杨海云[1] 肖越勇[2] LIU Gang;XIE Xiaoting;HE Hui;LIU Fei;MAO Xu;SANG Jingyao;YANG Haiyun;XIAO Yueyong(Department of Radiological Interventional Imaging,Qinghai Red Cross Hospital,Xining 810000,China;Department of Radiology,the First Medical Center,Chinese PLA General Hospital,Beijing 100853,China)

机构地区：[1]青海红十字医院放射影像介入科,青海西宁810000 [2]解放军总医院第一医学中心放射诊断科,北京100853

出　　处：《中国介入影像与治疗学》2024年第7期414-417,共4页Chinese Journal of Interventional Imaging and Therapy

基　　金：青海省基础研究计划-应用基础研究项目(2020-ZJ-781)。

摘　　要：目的 观察基于CT的残差神经网络(ResNet)-101-金字塔网络(FPN)模型鉴别肺良、恶性结节的价值。方法 回顾性分析2 000例肺结节患者共2 040个肺结节,包括良性1 150个、恶性890个;按8∶2比例将结节分为训练集(n=1 632)与测试集(n=408),前者包括良性结节881个、恶性结节751个,后者包括良性269个、恶性139个。以ResNet-101为主干网络、结合FPN基于胸部CT建立分类模型,观察其单一及联合医师评估鉴别肺良、恶性结节的效能。结果 测试集269个肺良性结节中,ResNet-101-FPN模型诊断正确214个(214/269,79.55%),联合医师评估后诊断正确230个(230/269,85.50%);139个恶性结节中,ResNet-101-FPN模型诊断正确124个(124/139,89.21%),联合医师评估后诊断正确131个(131/139,94.24%)。ResNet-101-FPN模型联合医师评估鉴别肺良、恶性结节的敏感度、准确率和精确度均高于而特异度则低于单一ResNet-101-FPN模型,但差异均无统计学意义(P均>0.05)。结论基于CT的ResNet-101-FPN模型可鉴别肺良、恶性结节;联合医师评估可提高诊断效能。Objective To observe the value of residual neural network(ResNet)-101-feature pyramid network(FPN)model based on CT for differentiating benign and malignant lung nodules.Methods Totally 2040 lung nodules in 2000 patients were retrospectively enrolled,including 1150 benign and 890 malignant nodules.The nodules were divided into training set(n=1632)and test set(n=408)at the ratio of 8∶2,the former including 881 benign and 751 malignant ones,while the latter including 269 benign and 139 malignant ones,respectively.Taken ResNet-101 as the backbone network,combined with FPN,a classification model was established based on chest CT,and the efficiency of this model alone and combined with evaluation of physicians for differentiating benign and malignant lung nodules were evaluated.Results Among 269 benign lung nodules in test set,ResNet-101-FPN model alone correctly diagnosed 214 nodules(214/269,79.55%),while combined with evaluation of physicians correctly diagnosed 230 ones(230/269,85.50%).For 139 malignant nodules in test set,ResNet-101-FPN model alone correctly diagnosed 124 nodules(124/139,89.21%),while combined with evaluation of physicians correctly diagnosed 131 ones(131/139,94.24%).The sensitivity,accuracy and precision of ResNet-101-FPN model combined with evaluation of physicians for distinguishing benign and malignant lung nodules were all higher,while the specificity of the combination was lower than those of ResNet-101-FPN model alone,but the differences were not significant(all P>0.05).Conclusion ResNet-101-FPN model could be used to distinguish benign and malignant lung nodules based on CT.Combining with evaluation of physicians could improve diagnostic efficiency of this model.

关 键 词：肺肿瘤 体层摄影术 X线计算机 人工智能 

分 类 号：R734.2[医药卫生—肿瘤] R814.42[医药卫生—临床医学]