摘要
目的:利用深度学习技术,全自动标注病变的计算机断层扫描(CT)数据,开发准确快速区分新型冠状病毒感染(COVID-19)和其他社区获得性肺炎的人工智能模型。方法:回顾性分析248例COVID-19患者及347例其他肺炎患者的资料,进行COVID-19与其他肺炎分类;在人工智能肺分割提取后将异常的CT图像特征降维,输入几种经典强化机器学习模型、三维卷积神经网络(3D CNN)和注意力多示例学习(Attention-MIL)深层神经网络架构中,模型诊断性能利用受试者工作特性(ROC)曲线、精确召回率(PR)曲线、曲线下面积(AUC)、敏感性、特异性、准确性指标进行评价。结果:在经典机器学习模型中K邻近算法(KNN)具有较好的效果,在外部测试集上的AUC值和平均精度(AP)值分别为0.79和0.89,平衡F分数(F1)值为0.76,准确率为0.75,敏感性为0.76,精确率为0.77;经典的3D CNN在外部测试集上效果良好,AUC值和AP值分别为0.64和0.82,F1值为0.71,准确率为0.78,敏感性为0.66,精确率为0.62;Attention-MIL模型在外部测试集上表现出更好的鲁棒性,AUC值和AP值分别为0.85和0.94,F1值达到0.82,准确率为0.92,敏感性为0.74,精确率为0.76。结论:与强化影像组学和3D CNN模型相比,深度学习Attention-MIL模型在鉴别诊断COVID-19和其他社区获得性肺炎上表现出更高的效能。
Purpose:Utilizing deep learning techniques,this study aimed to develop an artificial intelligence model that automatically annotates lesion computed tomography(CT)data,accurately and rapidly distinguishing novel coronavirus pneumonia(COVID-19)from other community-acquired pneumonia cases.Methods:A retrospective analysis was conducted on data from 248 patients with COVID-19 and 347 patients with other types of pneumonia.The COVID-19 cases were differentiated from other pneumonia cases during classification.After performing artificial intelligence-based lung segmentation,the extracted abnormal CT image features were dimensionally reduced and inputted into various classical machine learning models,Three-dimensional convolutional neural network(3D CNN),and attention-Multiple-instance learning(MIL)deep neural network architectures.The diagnostic performance of the models was evaluated using metrics such as receiver operating characteristic(ROC)curves,Precision Recall(PR)curves,Area Under Curve(AUC),sensitivity,specificity,and accuracy.Results:Among the classical machine learning models,K-Nearest Neighbor(KNN)demonstrated good performance,with an AUC of 0.793,Average Precision(AP)of 0.886,Balanced F Score(F1-score)of 0.7608,accuracy of 0.7512,sensitivity of 0.7754,and precision of 0.7691 on the external test set.The classical 3D CNN model exhibited satisfactory performance on the external test set with an AUC of 0.635,AP of 0.816,F1-score of 0.7144,accuracy of 0.7783,sensitivity of 0.6603,and precision of 0.6200.The attention-MIL model showed better robustness on the external test set,achieving an AUC of 0.851,AP of 0.935,F1-score of 0.8193,accuracy of 0.9155,sensitivity of 0.7414,and precision of 0.7646.Conclusion:Compared to the radiomics-enhanced and 3D CNN models,the deep learning attention-MIL model exhibited better performance in the differential diagnosis of COVID-19 and other community-acquired pneumonia.
作者
康兆庭
欧阳雪晖
柴军
KANG Zhaoting;OUYANG Xuehui;CHAI Jun(Department of Medical Imaging,Inner Mongolia Autonomous Region People's Hospital,Hohhot 010017,China)
出处
《CT理论与应用研究（中英文）》
2023年第5期685-694,共10页
Computerized Tomography Theory and Applications
基金
内蒙古自治区人民医院院内科研基金(基于机器学习弥漫低级别胶质瘤多参数MRI放射基因组学的病理分型与预后研究(2020YN17))
内蒙古自然科学基金(基于机器学习对弥漫低级别胶质瘤多参数MR放射基因组学的分子分型与预后研究(2021LHMS08066))
2022年度自治区医疗卫生科技计划项目(基于CT图像特征分析的机器学习算法预测非小细胞肺癌的病理分期与预后(202201038))。
