多模态人工智能模型对肺磨玻璃结节的良恶性的诊断研究  

Diagnosis of benign and malignant pulmonary ground-glass nodule by multimodal artificial intelligence model

作　　者：刘刚[1] 毛旭[1] 解晓婷 刘飞 何慧[1] 桑菁遥 杨海云[1] LIU Gang;MAO Xu;XIE Xiaoting;LIU Fei;HE Hui;SANG Jingyao;YANG Haiyun(Department of Radiology Interventional Imaging,Qinghai Red Cross Hospital,Xining 810000,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

出　　处：《实用放射学杂志》2023年第6期904-907,共4页Journal of Practical Radiology

基　　金：青海省基础研究计划-应用基础研究项目(2020-ZJ-781)。

摘　　要：目的建立基于人工智能的多模态模型对肺磨玻璃结节(GGN)良恶性诊断的研究。方法从各地区选取1020例患者按8︰2的比例分为训练集和测试集,从青海红十字医院选取204例患者作为验证集。所有患者临床数据包括人口统计学信息、肿瘤病史、肿瘤自身抗体、肿瘤指标以及病理和影像信息等。利用残差神经网络(ResNet)提取图像信息,Word2Vec方法提取语义信息以及Self-Attention方法融合图像和语义特征,最终构建多模态肺GGN良恶性分类模型,评估其准确度和敏感度。结果在测试集上多模态模型的准确度0.902(184/204)和敏感度0.966(197/204)均较计算机视觉、ResNet模型有良好的肺GGN良恶性诊断效能。在验证集中,多模态模型与影像科医生诊断正确的恶性GGN分别为125个(91.18%)和110个(80.37%),诊断正确的良性GGN分别为54个(80.70%)和57个(85.47%)。此外,影像科医生诊断结果、多模态模型诊断结果与病理结果的一致性检验提示,多模态模型诊断结果更接近金标准病理结果(Kappa=0.720,P<0.01)。结论人工智能多模态模型可高效准确地鉴别肺部GGN的良恶性,为临床医生的诊断提供可靠的辅助信息。Objective To establish a multimodal artificial intelligence model for the diagnosis of benign and malignant pulmonary ground-glass nodule(GGN).Methods A total of 1020 patients from various regions were collected,which were divided into training set and test set according to the ratio of 8︰2,and 204 patients from Qinghai Red Cross Hospital were collected as the validation set.All patients’clinical data included demographic information,tumor history,tumor autoantibodies,tumor indicators,pathological information and imaging information.Residual neural network(ResNet)was used to extract image information.Word2Vec method was used to extract semantic information and Self-Attention method was used to fuse image and semantic features.Thus,a multimodal classification model of benign and malignant pulmonary GGN was constructed to evaluate its accuracy and sensitivity.Results The accuracy 0.902(184/204)and sensitivity 0.966(197/204)of the multimodal model in the test set were better than those of the computer vision and the ResNet model in the diagnosis of benign and malignant pulmonary GGN.In the validation set,125(91.18%)and 110(80.37%)malignant GGN were correctly diagnosed by the multimodal model and the radiologists,while 54(80.70%)and 57(85.47%)benign GGN were correctly diagnosed,respectively.In addition,the consistency test of the diagnostic results of the radiologists,multimodal model and pathology indicated that the multimodal model were closer to the pathological results(Kappa=0.720,P<0.01).Conclusion The artificial intelligence multimodal model can accurately and efficiently identify the benign and malignant of pulmonary GGN,which can provide reliable supporting information for clinicians’diagnosis.

关 键 词：肺 磨玻璃结节 良恶性 人工智能 多模态 

分 类 号：R563[医药卫生—呼吸系统] TP183[医药卫生—内科学]