Textural features of ¹⁸F-fluorodeoxyglucose positron emission tomography scanning in diagnosing aortic prosthetic graft infection

For this retrospective analysis we included all patients (n = 30) from a prospective database of patients with a history of aortoiliac prosthetic graft reconstruction who underwent ¹⁸F-FDG PET/CT at our tertiary referral center between December 2009 and May 2015. Medical charts were analysed to identify those patients with a clinical suspicion of an AGI. An AGI was clinically suspected in case of undefined fever, a deep wound infection, persisting high laboratory infection parameters [e.g. erythrocyte sedimentation rate, white blood cell count, and/or C-reactive protein (CRP)], or a combination of these factors. AGI was considered proven only in cases where a positive bacterial culture of material was obtained from peri-prosthetic samples obtained during diagnostic work-up or surgery. Sixteen patients were clinically suspected to have an AGI (group I) of which 10 had a positive culture (group Ia) and six had a negative culture (group Ib). The remaining 14 patients underwent a ¹⁸F-FDG PET/CT scan because of cancer staging and were used as control group. Figure 1 shows a flow chart of the patient disposition.

Patient characteristics and information about the initial operation, type of graft material, clinical symptoms and laboratory parameters at the time of ¹⁸F-FDG PET imaging, and definite type of treatment were collected from the medical records. Co-morbidities were defined as recommended by the Ad Hoc Committee on Reporting Standards [24]. This study was approved by the institutional ethical review board (METc 2015/082). Patients’ data were analysed anonymously.

Non-gated PET/CT imaging was performed with a dedicated integrated PET/CT system (Biograph mCT PET/CT, Siemens, Knoxville, TN, USA). All patients fasted overnight with no restrictions on drinking water and with a minimum fasting time of 6 h prior to PET/CT. ¹⁸F-FDG was administered intravenously with a weight-based activity of 3 MBq/kg. Sixty minutes after tracer injection, patients were positioned on the camera table with the arms in upright position. PET images were acquired with 3 min per bed position. An initial low dose CT scan was performed to ensure that the region of interest was included in the field of view, where after an inspiration breath-hold low-dose CT for attenuation correction was performed with 100 kVp and 30 mAs. Image data were reconstructed using standard methods and images were standardized according to EANM guidelines [13].

An experienced nuclear medicine physician assessed the ¹⁸F-FDG PET images, including VGS, SUVmax and TBR. The five-point VGS was graded as follows: grade 0, ¹⁸F-FDG uptake similar to that in the background; grade I, low ¹⁸F-FDG uptake, comparable with inactive muscles and fat; grade II, moderate ¹⁸F-FDG uptake, clearly visible and higher than uptake by inactive muscles and fat; grade III, strong ¹⁸F-FDG uptake, but distinctly less than the physiologic urine bladder activity; and grade IV, very strong ¹⁸F-FDG uptake, comparable with the physiologic urinary activity of the bladder [10, 25]. A VOI was drawn around the area of the vascular prosthesis to calculate the SUVmax. SUVmax corresponded to the voxel with the highest ¹⁸F-FDG uptake. The TBR was defined as the SUVmax divided by the mean SUV of the caval vein (blood pool).

¹⁸F-FDG PET based textural features to characterize heterogeneity of ¹⁸F-FDG uptake in the aortic prosthetic graft were measured for each patient. Figure 2 displays the image processing and extraction of all textural features (n = 64) from a predefined VOI. VOIs were manually delineated (Fig. 2a) by two independent experienced nuclear medicine physicians on axial planes of the low-dose CT to enclose three-dimensional coverage of the entire suspected prosthetic graft, using PMOD 3.6 software (PMOD Inc., Zurich, Switzerland). However, as the partial volume effect may cause the activity to be smeared out over a larger area than the actual structure, and the total number of counts is preserved, we also included high uptake regions around the graft material with clear contamination of activity from the graft (spill over). Moreover, we assessed whether ¹⁸F-FDG uptake was physiological or non-physiological based on the CT-based anatomical location and excluded high uptake regions with clear spill-in effects from physiological neighbouring tissues such as the kidneys, ureter, and urine bladder. Although this VOI definition leads to analysis on the whole prosthetic graft volume, we chose for this definition since analysis of the most-diseased segment of the aortic graft would require a semi-objective identification of the ¹⁸F-FDG-avid area [13]. Additionally, prosthetic grafts of group II do often not contain a ¹⁸F-FDG-avid area, which, therefore, complicates the comparison with Group I. After VOI delineation, the PET/CT imaging data and VOI delineations were loaded into Matlab 2014b (Mathworks, Natick, MA, USA) for processing and analysis (Fig. 2b). The VOI delineations were then registered to the PET images. PET voxels, which were enclosed for ≥50% coverage, were considered to be part of the suspected prosthetic graft (Fig. 2c). For noise reduction, SUV was discretized with fixed 0.5 g/mL increments according to Doane’s optimal bin width [26]. Textural analysis was performed on the cropped PET VOI. To find the influence of the volume of the VOI on textural features the Pearson correlation coefficient was calculated.

Figure 2d and supplemental table 1 provide a full overview of all 64 analysed textural features. We extracted 19 first order textural features (based on the grey level distribution, but without spatial information of voxels). Texture can be characterized by replications of (small) texture elements. These texture elements consist of contiguous voxels with certain spatial and intensity properties. We obtained the distributions of three different texture elements, i.e. the grey level co-occurrence (or spatial dependence) matrix (GLCM) for pairwise arrangement of voxels [27], the grey level run-length matrix (GLRLM) for alignment of voxels with an identical intensity [28], and the grey level size-zone matrix (GLSZM) for characteristics of homogeneous zones [29]. From these matrices, we extracted 46 s order textural features (which are thus based on spatial information of the grey levels). These textural features were extracted with a voxel-to-voxel distance offset of d = 1 and directional analysis was performed with a connectivity of 26 voxels (analysis in 13 angular directions). All extracted textural features were normalized to the range [0,1]. To determine the influence of noise, it was tested whether noise was equally distributed among the groups and the correlation between noise and each textural feature was computed. Therefore, a sphere of 3 cm in diameter was drawn in the liver; the coefficient of variation was determined as noise parameter.

Baseline characteristics are presented as mean ± standard deviation or percentages. To select candidate predictors to identify infection, univariable logistic regression analysis was performed. All potential predictors that met the Akaike Information Criterion (AIC) were considered significant [30]. To discourage overfitting, the AIC is based on rewarding goodness of fit and penalizing complexity in the model. The AIC requires χ ² > 2 df, i.e. when considering a predictor with one degree of freedom df; this implies a significance level α = P(χ ² ≥ 2) = 0.16 [31]. The accuracy of all candidate predictors was measured by the area under the receiver operating characteristic curve (AUC). Textural features were considered to have a good accuracy in case of an AUC > 0.8. Moreover, textural feature values may be subject to inter-observer variability in delineation of the prosthesis. Textural features were considered stable in case of a minimum acceptable excellent agreement indicated by an intra-class correlation coefficient (ICC) level of 0.75 [32]. To obviate multicollinearity among all significant, accurate, and stable considered textural features, the pairwise Pearson correlation coefficient was evaluated. When the correlation of a pair of variables was >0.8, the variable with the lowest AUC was excluded from the set of features chosen for AGI characterization. Data were collected and analysed using IBM SPSS 20.0 software (IBM Corp, Armonk, NY, USA).