Montgomery Sean, Li Forrest, Funk Christopher, Peethumangsin Erica, Morris Michael, Anderson Jess T, Hersh Andrew M, Aylward Stephen
From the Duke University Hospital (S.M., E.P.), Durham, North Carolina; Kitware, Inc (F.L., C.F., S.A.), Carrboro, North Carolina; Brooke Army Medical Center (M.M., J.T.A.), San Antonio, Texas; and Montrose Regional Health (A.M.H.), Montrose, Colorado.
J Trauma Acute Care Surg. 2023 Mar 1;94(3):379-384. doi: 10.1097/TA.0000000000003845. Epub 2022 Nov 28.
Ultrasound (US) for the detection of pneumothorax shows excellent sensitivity in the hands of skilled providers. Artificial intelligence may facilitate the movement of US for pneumothorax into the prehospital setting. The large amount of training data required for conventional neural network methodologies has limited their use in US so far.
A limited training database was supplied by Defense Advanced Research Projects Agency of 30 patients, 15 cases with pneumothorax and 15 cases without. There were two US videos per patient, of which we were allowed to choose one to train on, so that a limited set of 30 videos were used. Images were annotated for ribs and pleural interface. The software performed anatomic reconstruction to identify the region of interest bounding the pleura. Three neural networks were created to analyze images on a pixel-by-pixel fashion with direct voting determining the outcome. Independent verification and validation was performed on a data set gathered by the Department of Defense.
Anatomic reconstruction with the identification of ribs and pleura was able to be accomplished on all images. On independent verification and validation against the Department of Defense testing data, our program concurred with the SME 80% of the time and achieved a 86% sensitivity (18/21) for pneumothorax and a 75% specificity for the absence of pneumothorax (18/24). Some of the mistakes by our artificial intelligence can be explained by chest wall motion, hepatization of the underlying lung, or being equivocal cases.
Using learning with limited labeling techniques, pneumothorax was identified on US with an accuracy of 80%. Several potential improvements are controlling for chest wall motion and the use of longer videos.
Diagnostic Tests; Level III.
超声（US）用于检测气胸在技术熟练的操作者手中显示出极高的灵敏度。人工智能可能有助于将用于检测气胸的超声应用于院前环境。传统神经网络方法所需的大量训练数据限制了其目前在超声领域的应用。
由美国国防高级研究计划局提供了一个有限的训练数据库，包含30例患者，其中15例气胸患者和15例非气胸患者。每位患者有两段超声视频，我们被允许选择其中一段用于训练，因此总共使用了30段有限的视频。对肋骨和胸膜界面进行图像标注。该软件进行解剖重建以识别界定胸膜的感兴趣区域。创建了三个神经网络，以逐像素方式分析图像，并通过直接投票确定结果。对美国国防部收集的数据集进行了独立验证。
所有图像均能完成肋骨和胸膜的解剖重建及识别。在针对美国国防部测试数据的独立验证中，我们的程序在80%的情况下与专家意见一致，对气胸的检测灵敏度达到86%（21例中的18例），对无气胸情况的特异性为75%（24例中的18例）。我们的人工智能出现的一些错误可以通过胸壁运动、肺实质肝样变或病例不明确来解释。
使用有限标注技术进行学习，超声检测气胸的准确率为80%。几个潜在的改进方向是控制胸壁运动以及使用更长的视频。
诊断试验；三级。
