From the Department of Radiology, University Hospital Basel, Basel, Switzerland.
Invest Radiol. 2019 Jan;54(1):55-59. doi: 10.1097/RLI.0000000000000509.
The aim of this study was to test the diagnostic performance of a deep learning-based triage system for the detection of acute findings in abdominal computed tomography (CT) examinations.
Using a RIS/PACS (Radiology Information System/Picture Archiving and Communication System) search engine, we obtained 100 consecutive abdominal CTs with at least one of the following findings: free-gas, free-fluid, or fat-stranding and 100 control cases with absence of these findings. The CT data were analyzed using a convolutional neural network algorithm previously trained for detection of these findings on an independent sample. The validation of the results was performed on a Web-based feedback system by a radiologist with 1 year of experience in abdominal imaging without prior knowledge of image findings through both visual confirmation and comparison with the clinically approved, written report as the standard of reference. All cases were included in the final analysis, except those in which the whole dataset could not be processed by the detection software. Measures of diagnostic accuracy were then calculated.
A total of 194 cases were included in the analysis, 6 excluded because of technical problems during the extraction of the DICOM datasets from the local PACS. Overall, the algorithm achieved a 93% sensitivity (91/98, 7 false-negative) and 97% specificity (93/96, 3 false-positive) in the detection of acute abdominal findings. Intra-abdominal free gas was detected with a 92% sensitivity (54/59) and 93% specificity (39/42), free fluid with a 85% sensitivity (68/80) and 95% specificity (20/21), and fat stranding with a 81% sensitivity (42/50) and 98% specificity (48/49). False-positive results were due to streak artifacts, partial volume effects, and a misidentification of a diverticulum (each n = 1).
The algorithm's autonomous detection of acute pathological abdominal findings demonstrated a high diagnostic performance, enabling guidance of the radiology workflow toward prioritization of abdominal CT examinations with acute conditions.
本研究旨在测试基于深度学习的分诊系统在检测腹部计算机断层扫描（CT）检查中急性发现方面的诊断性能。
使用 RIS/PACS（放射信息系统/图片存档和通信系统）搜索引擎，我们获得了 100 例连续的腹部 CT 检查，这些 CT 检查至少有以下一种发现：游离气体、游离液体或脂肪线，并获得了 100 例无这些发现的对照病例。使用之前在独立样本上针对这些发现进行训练的卷积神经网络算法对 CT 数据进行分析。结果通过一位具有 1 年腹部成像经验的放射科医生在基于网络的反馈系统上进行验证，该医生在没有事先了解图像发现的情况下通过视觉确认和与临床批准的书面报告进行比较来作为参考标准。除了由于从本地 PACS 中提取 DICOM 数据集时出现技术问题而无法处理整个数据集的 6 例病例外，所有病例均纳入最终分析。然后计算诊断准确性的度量。
共有 194 例病例纳入分析，6 例病例由于从本地 PACS 中提取 DICOM 数据集时出现技术问题而被排除。总的来说，该算法在检测急性腹部发现方面的灵敏度为 93%（91/98，7 例假阴性），特异性为 97%（93/96，3 例假阳性）。腹腔内游离气体的检测灵敏度为 92%（54/59），特异性为 93%（39/42）；游离液体的检测灵敏度为 85%（68/80），特异性为 95%（20/21）；脂肪线的检测灵敏度为 81%（42/50），特异性为 98%（48/49）。假阳性结果是由于条纹伪影、部分容积效应和憩室的误识别（各 n=1）。
该算法自主检测急性病理性腹部发现的性能表现出很高的诊断性能，可指导放射科工作流程，优先进行有急性情况的腹部 CT 检查。
