Artificial Intelligence Augmentation of Radiologist Performance in Distinguishing COVID-19 from Pneumonia of Other Origin at Chest CT
Artificial Intelligence Augmentation of Radiologist Performance in Distinguishing COVID-19 from Pneumonia of Other Origin at Chest CT
复制标题
人工智能增强放射科医生在胸部 CT 区分 COVID-19 与其他来源的肺炎方面的表现
DOI：
10.1148/radiol.2020201491
复制
发表时间：
2020-09-01
期刊：
影响因子：
19.7
通讯作者：
Liao, Wei-Hua
中科院分区：
文献类型：
--
作者：
Bai, Harrison X.;Wang, Robin;Liao, Wei-Hua
背景资料：2019冠状病毒病新冠肺炎（COVID-19）和其他疾病的肺炎具有相似的CT特征，这有助于高准确度区分它们的挑战。目的：建立和评估人工智能（AI）系统，用于在胸部CT上区分COVID-19和其他肺炎，并评估放射科医生在没有和有AI辅助的情况下的表现。材料和方法：从2020年1月至2020年4月，回顾性地从10家医院中识别出521例COVID-19逆转录聚合酶链反应结果阳性且胸部CT结果异常的患者。回顾性选取了2017年至2019年三家医院共665例非COVID-19肺炎且胸部CT有明确肺炎证据的患者。为了对每位患者的COVID-19与其他肺炎进行分类，在肺部分割后，将异常CT切片输入到Efficient Net B4深度神经网络架构中，然后由两层全连接神经网络将切片汇集在一起。最终的1186例患者（132583个CT切片）按7：2：1等比例分为训练集、验证集和测试集。通过在不同医院评价模型性能进行独立测试。研究由六名放射科医生在没有AI辅助的情况下进行盲审，然后在AI辅助下进行盲审。结果：最终模型的测试准确率达到96%（95%置信区间[CI]：90%，98%），灵敏度为95%（95% CI：83%，100%），特异性为96%（95% CI：88%，99%），受试者操作特征曲线下面积为0.95，精确-召回曲线下面积为0.90。在独立测试中，该模型的准确性为87%（95% CI：82%，90%），灵敏度为89%（95% CI：81%，94%），特异性为86%（95% CI：80%，90%），受试者工作特征曲线下面积为0.90，精确度-召回率曲线下面积为0.87。在模型概率的辅助下，放射科医师获得了更高的平均测试准确率（90%对85%，Delta= 5，P
Background: Coronavirus disease 2019 (COVID-19) and pneumonia of other diseases share similar CT characteristics, which contributes to the challenges in differentiating them with high accuracy.Purpose: To establish and evaluate an artificial intelligence (AI) system for differentiating COVID-19 and other pneumonia at chest CT and assessing radiologist performance without and with AI assistance.Materials and Methods: A total of 521 patients with positive reverse transcription polymerase chain reaction results for COVID-19 and abnormal chest CT findings were retrospectively identified from 10 hospitals from January 2020 to April 2020. A total of 665 patients with non-COVID-19 pneumonia and definite evidence of pneumonia at chest CT were retrospectively selected from three hospitals between 2017 and 2019. To classify COVID-19 versus other pneumonia for each patient, abnormal CT slices were input into the Efficient Net B4 deep neural network architecture after lung segmentation, followed by a two-layer fully connected neural network to pool slices together. The final cohort of 1186 patients (132 583 CT slices) was divided into training, validation, and test sets in a 7:2:1 and equal ratio. Independent testing was performed by evaluating model performance in separate hospitals. Studies were blindly reviewed by six radiologists without and then with AI assistance.Results: The final model achieved a test accuracy of 96% (95% confidence interval [CI]: 90%, 98%), a sensitivity of 95% (95% CI: 83%, 100%), and a specificity of 96% (95% CI: 88%, 99%) with area under the receiver operating characteristic curve of 0.95 and area under the precision-recall curve of 0.90. On independent testing, this model achieved an accuracy of 87% (95% CI: 82%, 90%), a sensitivity of 89% (95% CI: 81%, 94%), and a specificity of 86% (95% CI: 80%, 90%) with area under the receiver operating characteristic curve of 0.90 and area under the precision-recall curve of 0.87. Assisted by the probabilities of the model, the radiologists achieved a higher average test accuracy (90% vs 85%, Delta= 5, P