Cerebral PCT involves repeated sequential image acquisition of the head to obtain dynamic CT images of the brain parenchyma following intravenous administration of an iodinated contrast bolus. Data are postprocessed by vendor’s workstations to generate parametric perfusion color maps that can be evaluated qualitatively and quantitatively [1].

For acute infarction as an example, areas of irreversibly infarcted core show decreased CBF and CBV but increased MTT, while areas of potentially salvageable penumbra show decreased CBF with normal CBV but prolonged MTT indicating a CBV/MTT mismatch [2]. However, the calculation accuracy can be affected by variations in image acquisition protocols and image postprocessing methods. Image acquisition parameters such as sampling times and imaging duration which depend on the scanners used might be different among institutions. There are also different postprocessing algorithms used by vendors to generate perfusion maps, such as singular-value decomposition, inverse filter, maximum slope and box modular transfer function, etc [3]. Broadly speaking, they can be classified into deconvolution and nondeconvolution methods. Another variation in image postprocessing method would be whether the arterial input function (AIF) and venous outflow function (VOF) are selected manually or automatically by the software. The results of manual placement might be operator dependent, while automated placement might introduce errors due to inappropriate selection of the vascular structure or volume averaging.

A CT perfusion phantom (Sun Nuclear Corporation, Melbourne, FL) configured with a travel time of 70 ± 1.4 s was scanned on Revolution CT (GE Healthcare, Chicago, IL, USA) in axial mode (Fig. 1). The acquisition parameters were 80 kVp, 200 mA, 40 mm detector collimation, 5 mm slice thickness, 0.5 s rotation time, 40 passes, temporal resolution of 2 s and scan duration of 20 s. The PCT dataset was transferred to workstations installed with (A) CT Neuro Perfusion workflow, syngo.via, VB60A_HF05 (Siemens Healthineers, Erlangen, Germany); and (B) CT Perfusion 4D application, AW Server 3.2 Ext. 4.2 (GE Healthcare, Chicago, IL, USA). Same source dataset was used to avoid variations in image acquisition protocols and ensure same source data quality. Delay-insensitive deconvolution algorithm was selected in package A, whereas it was the fixed algorithm in package B. The central image slice was selected for postprocessing in both packages. For package A, stroke template was used with brain segmentation deactivated and 4D noise reduction activated. The 4D noise reduction algorithm uses a spatiotemporal multi-band filtering approach. In the dynamic series, images from every time frame are decomposed into multiple frequency bands. Different weighting functions are applied for different frequency bands that are recombined to form the final image. As a result, there will be fewer areas of the perfusion maps where the postprocessing algorithm fails due to insufficient signal-to-noise ratio. For package B, CT Brain Stroke protocol was used with cranium removal and vessel exclusion disabled. AIF and VOF were selected manually. Time-attenuation curves (TACs) of AIF, VOF and two brain tissue rods were plotted in the two packages to derive the quantitative perfusion parameters (Fig. 2 and Fig. 3). The small regions of interest to select artery and vein in package B resulted in some noise in their TACs. Circular ROIs were drawn to circumscribe the two brain tissue rods on the parametric perfusion color maps (Fig. 4 and Fig. 5). Mean values and standard deviations of CBF, CBV and MTT for each ROI were compared between the two packages. The whole process was done by the same operator for ten times to minimize the intraoperator variability.