基于CT深度学习的活动性/非活动性肺结核分类模型构建及验证

Establishment and verification of active/non-active tuberculosisclassification model based on CT deep learning

目的:评价基于深度学习框架(3D ResNet-50)的肺结核CT辅助诊断模型在活动性/非活动性肺结核鉴别诊断中的临床应用价值。方法:回顾性收集2018年1月至2020年12月期间喀什地区第一人民医院经结核分枝杆菌培养检测并接受胸部高分辨率CT平扫检查的1 940例患者的病例资料,活动性肺结核患者(ATB)960例,非活动性肺结核患者(non-ATB)980例。所有数据以7∶2∶1的比例随机分组,用于训练、验证和测试集,使用3D Nested UNet模型预先分割肺野区域,然后使用3D ResNet-50进行分类,并收集了用于模型外部验证的外部测试集ATB 100例,non-ATB 100例。为了量化评估深度学习模型的性能,绘制ROC曲线,以曲线下面积(AUC)、准确率(ACC)、召回率(Recall)、F1分数(F1 score)作为模型评价指标,并使用类激活图来评估感兴趣的激活区域,最后,将该模型与两位放射科医生的外部测试数据进行了比较。结果:本组ATB患者比非ATB患者年龄大。结果显示ATB患者咳嗽更多,而非ATB患者出现更多胸痛症状,差异有统计学意义(P<0.05)。模型对训练集、验证集、测试集及外部验证的诊断性能AUC分别为0.970、0.951、0.933、0.942;ACC分别为0.996、0.981、0.975、0.967,Recall分别为0.989、0.953、0.951、0.952,F1 score分别为0.995、0.983、0.972、0.969。此外,3D ResNet-50模型的性能高于两名放射科医生,AI模型诊断效能也比放射科医师快10倍。结论:3D ResNet-50 AI系统诊断水平与经验丰富的放射科医生水平接近,可作为活动性肺结核检测及鉴别诊断的快速辅助诊断工具,快速区分活动性和非活动性肺结核。

Objective:To evaluate the clinical application value of CT assisted diagnosis model of pulmonary tuberculosis based on deep learning framework(3D ResNet-50)in the differential diagnosis of active/non-active tuberculosis.Methods:The case data of 1940 patients who were tested by Mycobacterium tuberculosis culture and underwent high-resolution CT scan of the chest at the First People s Hospital of Kashgar Region during January 2018 to December 2020 were retrospectively collected,960 patients with active pulmonary tuberculosis(ATB)and 980 patients with non-active pulmonary tuberculosis(non-ATB).All data were randomized in a 7∶2∶1 ratio for training,verification,and test sets,pre-segmented lung field areas using a 3D Nested UNet model,followed by classification using 3D ResNet-50,and 100 external test sets ATB and 100 non-ATBs were collected for external validation of the model.In order to quantify the performance of the deep learning model,ROC curves were plotted with area under the curve(AUC),accuracy rate(ACC),recall rate(Recall)and F1 score(F1 score)as model evaluation indicators,and class activation plots were used to evaluate the activation region of interest.Finally,the model was compared with external test data from two radiologists.Results:ATB patients were older than non-ATB patients.The results show,more coughing was observed in ATB patients,and more chest pain was experienced by non-ATB patients;the difference was statistically significant(P<0.05).The diagnostic performances of the model for training set,validation set,test set and external validation were AUC=0.970,0.951,0.9330.942;ACC=0.996,0.981,0.975,0.967,Recall were 0.989,0.953,0.951,0.952 respectively,and F1 score was 0.995,0.983,0.972,0.969 respectively.In addition,the 3D ResNet-50 model performed better than two radiologists and the AI model was 10 times faster than the radiologist.Conclusion:The 3D ResNet-50 AI system is diagnostic at a level close to that of an experienced radiologist and can be used as a rapid diagnostic aid for the detection and differential diagnosis of active tuberculosis,quickly distinguishing between active and non-active tuberculosis.