Can Deep Learning Reliably Recognize Abnormality Patterns on Chest X-rays? A Multi-Reader Study Examining One Month of AI Implementation in Everyday Radiology Clinical Practice

Learning · Seven · 假阴性 · DSS · Performer ·

2023 年 6 月 2 日

翻译：深度学习能否可靠识别胸部X光片的异常模式？一项为期一个月的人工智能临床实施多读片者研究

Daniel Kvak,Anna Chromcová,Petra Ovesná,Jakub Dandár,Marek Biroš,Robert Hrubý,Daniel Dufek,Marija Pajdaković

In this study, we developed a deep-learning-based automatic detection algorithm (DLAD, Carebot AI CXR) to detect and localize seven specific radiological findings (atelectasis (ATE), consolidation (CON), pleural effusion (EFF), pulmonary lesion (LES), subcutaneous emphysema (SCE), cardiomegaly (CMG), pneumothorax (PNO)) on chest X-rays (CXR). We collected 956 CXRs and compared the performance of the DLAD with that of six individual radiologists who assessed the images in a hospital setting. The proposed DLAD achieved high sensitivity (ATE 1.000 (0.624-1.000), CON 0.864 (0.671-0.956), EFF 0.953 (0.887-0.983), LES 0.905 (0.715-0.978), SCE 1.000 (0.366-1.000), CMG 0.837 (0.711-0.917), PNO 0.875 (0.538-0.986)), even when compared to the radiologists (LOWEST: ATE 0.000 (0.000-0.376), CON 0.182 (0.070-0.382), EFF 0.400 (0.302-0.506), LES 0.238 (0.103-0.448), SCE 0.000 (0.000-0.634), CMG 0.347 (0.228-0.486), PNO 0.375 (0.134-0.691), HIGHEST: ATE 1.000 (0.624-1.000), CON 0.864 (0.671-0.956), EFF 0.953 (0.887-0.983), LES 0.667 (0.456-0.830), SCE 1.000 (0.366-1.000), CMG 0.980 (0.896-0.999), PNO 0.875 (0.538-0.986)). The findings of the study demonstrate that the suggested DLAD holds potential for integration into everyday clinical practice as a decision support system, effectively mitigating the false negative rate associated with junior and intermediate radiologists.

翻译：本研究开发了一种基于深度学习的自动检测算法（DLAD, Carebot AI CXR），用于检测并定位胸部X光片（CXR）中七种特定放射学征象（肺不张（ATE）、实变（CON）、胸腔积液（EFF）、肺结节（LES）、皮下气肿（SCE）、心脏增大（CMG）、气胸（PNO））。我们收集了956张胸部X光片，并比较了该DLAD与六名放射科医生在医院环境下评估图像时的表现。所提出的DLAD达到了较高的灵敏度（肺不张1.000（0.624-1.000）、实变0.864（0.671-0.956）、胸腔积液0.953（0.887-0.983）、肺结节0.905（0.715-0.978）、皮下气肿1.000（0.366-1.000）、心脏增大0.837（0.711-0.917）、气胸0.875（0.538-0.986）），即便与放射科医生相比亦表现优异（最低值：肺不张0.000（0.000-0.376）、实变0.182（0.070-0.382）、胸腔积液0.400（0.302-0.506）、肺结节0.238（0.103-0.448）、皮下气肿0.000（0.000-0.634）、心脏增大0.347（0.228-0.486）、气胸0.375（0.134-0.691）；最高值：肺不张1.000（0.624-1.000）、实变0.864（0.671-0.956）、胸腔积液0.953（0.887-0.983）、肺结节0.667（0.456-0.830）、皮下气肿1.000（0.366-1.000）、心脏增大0.980（0.896-0.999）、气胸0.875（0.538-0.986））。本研究结果显示，所提出的DLAD作为决策支持系统具有整合至日常临床实践的潜力，能够有效降低中初级放射科医生的假阴性率。