Große Hokamp Nils, Eck Brendan, Siedek Florian, Pinto Dos Santos Daniel, Holz Jasmin A, Maintz David, Haneder Stefan
Institute for Diagnostic and Interventional Radiology, Faculty of Medicine and University Hospital Cologne, University of Cologne, Cologne, Germany.
Department of Radiology, Case Western Reserve University, Cleveland, OH, USA.
Quant Imaging Med Surg. 2020 May;10(5):1033-1044. doi: 10.21037/qims.2020.04.03.
Numerous methods for artifact quantification in computed tomography (CT) imaging have been suggested. This study evaluated their utility with regards to correspondence with visual artifact perception and reproducibility. Two titanium rods (5 and 10 mm) were examined with 25 different scanning- and image-reconstruction parameters resulting in different types and extents of artifacts. Four radiologists evaluated every image against each other using an in-house developed software. Rating was repeated two times (2,400 comparisons = 2 times × 4 readers × 300 comparisons). Rankings were combined to obtain a reference ranking. Proposed approaches for artifact quantification include manual measurement of attenuation, standard deviation and noise and sophisticated algorithm-based approaches within the image- and frequency-domain. Two radiologists conducted manual measurements twice while the aforementioned algorithms were implemented within the Matlab-Environment allowing for automated image analysis. The reference ranking was compared to all aforementioned methods for artifact quantification to identify suited approaches. Besides visual analysis, Kappa-statistics and intraclass correlation coefficients (ICC) were used. Intra- and Inter-reader agreements of visual artifact perception were excellent (ICC 0.85-0.92). No quantitative method was able to represent the exact ranking of visually perceived artifacts; however, ICC for manual measurements were low (ICC 0.25-0.97). The method that showed best correspondence and reproducibility used a Fourier-transformed linear ROI and lower-end frequency bins. Automated measurements of artifact extent should be preferred over manual measurements as the latter show a limited reproducibility. One method that allows for automated quantification of such artefacts is made available as an electronic supplement.
人们已经提出了许多用于计算机断层扫描（CT）成像中伪影量化的方法。本研究评估了它们在与视觉伪影感知的对应性和可重复性方面的效用。使用25种不同的扫描和图像重建参数对两根钛棒（5毫米和10毫米）进行了检查，这些参数导致了不同类型和程度的伪影。四名放射科医生使用内部开发的软件相互评估每幅图像。评级重复进行了两次（2400次比较 = 2次×4名读者×300次比较）。将排名合并以获得参考排名。提出的伪影量化方法包括手动测量衰减、标准差和噪声，以及在图像域和频率域中基于复杂算法的方法。两名放射科医生进行了两次手动测量，同时在Matlab环境中实现了上述算法，以实现自动图像分析。将参考排名与所有上述伪影量化方法进行比较，以确定合适的方法。除了视觉分析外，还使用了kappa统计和组内相关系数（ICC）。视觉伪影感知的读者内和读者间一致性非常好（ICC 0.85 - 0.92）。没有一种定量方法能够准确代表视觉感知伪影的排名；然而，手动测量的ICC较低（ICC 0.25 - 0.97）。显示出最佳对应性和可重复性的方法使用了傅里叶变换线性感兴趣区域和低端频率区间。与手动测量相比，应优先选择伪影范围的自动测量，因为后者的可重复性有限。一种允许自动量化此类伪影的方法作为电子补充材料提供。
