Table of Contents
Patients
Ethical permission was obtained from the medical ethical committee Arnhem-Nijmegen (ClinicalTrials.gov identifier: NCT04596943). All study proceedings were performed in accordance with Dutch clinical trials guidelines and all participants provided written informed consent prior to participation. In this single-center proof-of-concept prospective observational study, we included hospitalized adult patients with PCR proven SARS-CoV-2 infection, admitted to the nursing ward. Exclusion criteria included previously documented severe lung abnormalities, glomerular filtration rate ≤ 30 ml/min, contra-indications for PET/CT (pregnancy, breast-feeding or severe claustrophobia) or contra-indications for administration of iodine-containing agents. Patient data, including demographics, medical history, clinical parameters, laboratory examinations, treatment and complications during hospital stay were collected. No adverse events were reported.
Non-COVID-19 patients
Five patients with oral cavity squamous cell carcinomas were used as reference. 68Ga-RGD PET/CT scans were performed according to the protocol described in the 2016 study of Lobeek et al.22.
Image acquisition
68Ga-RGD PET/CT
[68Ga]Ga-DOTA-E-[c(RGDfK)]2 was synthesized at the Radboudumc (Nijmegen, the Netherlands) as described in Lobeek et al.22 A mean dose of 196 ± 20 MBq 68Ga-RGD was injected intravenously as a bolus over 1 min followed by saline flushing. All PET/CT scans were performed on a Biograph mCT 4-ring clinical scanner without ECG or respiratory gating (Siemens). PET acquisition of patients 2–10 commenced median 31 min (IQR 24–38) post-injection at 5 min per bed position. Patient 1 was scanned at a significantly later time point post-injection (118 min) due to patient transport logistics, this subject showed a remarkably lower 68Ga-RGD uptake across all analyses. The scan range included the thorax, head and neck of the patients. Reconstruction of PET images with vendor specific software comprised of attenuation correction with CT and TrueX algorithm with point spread function and time of flight measurement using 3 iterations and 21 subsets (Siemens). Slice thickness was 3 mm, pixel spacing 4.07 mm, matrix size 200 × 200 voxels and pixel full-width half maximum 3 mm. A 3D Gaussian filter kernel of 3 mm was used for postprocessing.
Low-dose CT scans for attenuation correction (CTac) and anatomical reference were acquired with automatically modulated X-ray tube voltage and current (120 kV, 50 mA). Scan range was equal to PET, slice thickness 3 mm, pixel spacing 0.98 mm, matrix size 512 × 512 voxels and images were reconstructed using a B31f kernel.
Subtraction CT
Directly following PET/CT, patients underwent subtraction CT on the same scanner: A CT of the thorax before and after iodinated intravenous contrast administration (iomeprol 300mgl/ml), to evaluate one-phase iodine enhancement of the pulmonary parenchyma. An unenhanced CT (CTld, mean DLP 124 mGy.cm) was made after breath hold instruction with automatically modulated X-ray tube current (reference 75mAs, > 66). Subsequently, injection of a bolus (112 ± 12 ml) of 300 mg/ml iodine-contrast at 5 ml/s was followed by a 40 ml saline chaser at the same injection rate. One patient received a contrast bolus of 3 ml/s and consequently a triggering delay. After a threshold of 100 hounsfield units (HU) was measured in the pulmonary trunk, a breath hold instruction was given to the patient for the acquisition of CT angiography (CTA) of the thorax using modulated current (reference 100 mAs, > 86, mean DLP = 169 mGy cm). Both CT images were acquired with tube voltage of 100 kV and reconstructed with kernel I30f/3, a slice thickness of 1.5 mm with pixel spacing of 0.72 mm and matrix voxel size 512 × 512. Median HU in de pulmonary trunk was 414 HU.
From these two scans, iodine maps were calculated by subtracting Ctld from the CTA scans after motion correction and mask segmentation as described in Grob et al.23.
Image analysis
Lung segmentation
We used a previously developed COVID-CT artificial intelligence algorithm for segmentation of the five lung lobes in the CTld images24. Additionally, this algorithm segmented affected areas (with GGOs and consolidations) from unaffected areas per lobe. This resulted in 10 ROIs per patient, and the corresponding CT severity score per lung lobe as described in Lessmann et al.24 The CTSS was not calculated for the reference group, since this score is validated for COVID-19 and not for COPD associated changes in lung parenchyma. Rigid registration of the CTld to the CTac was performed using MevisLab (Fraunhofer Mevis, Bremen, Germany). This transformation was used to register the segmentation to the PET images and subsequently calculate the SUV within the 10 ROIs.
As the PET scan was made during free breathing, one investigator manually adjusted the segmentations to exclude the liver and spleen signal from quantification if their activity concentration was projected over the lower lung.
68Ga-RGD PET/CT scans of the reference patients were used as reference PET signal in lung parenchyma unaffected by COVID-19 infection. The CTac images of the references were segmented using the same lung lobe segmentation artificial intelligence algorithm24. This segmentation was registered to the PET images and subsequently adjusted to exclude the liver and spleen signal before calculating mean SUV per lobe.
Subtraction CT
One investigator (EvG) and one chest radiologist with 6 years of experience in thoracic radiology (MB) evaluated image quality and presence and grade of perfusion inhomogeneities on subtraction CT. They graded image quality of the perfusion maps per lung lobe on a visual grading scale from 1 to 3 (1: bad, 2: acceptable, 3: good). In each of the 10 ROIs per patient perfusion was assessed on a scale from 1 to 5 (1: severely decreased, 2: decreased, 3: as expected, 4: increased, 5: severely increased) compared to what was expected in a corresponding part of healthy parenchyma. In case of discrepancy between the two readers, this was solved in consensus.
Myocardium segmentation
The myocardium of the left ventricle (LV) was delineated on the CTld using the artificial intelligence based algorithm “Whole-heart segmentation in non-contrast-enhanced CT” in all patients44. On basis of this segmentation, the mean SUV was calculated for the myocardium. Additionally, one investigator manually (FVDH) delineated the LV myocardium on the CTA of the COVID-19 patients. The mean SUV derived from the algorithm was compared per patient with the mean SUV derived from the manual delineation in order to verify the algorithm. The relatively thin wall of the right ventricle was not delineated as SUV quantification would be unreliable on a scan without ECG or respiratory gating.
Carotid artery segmentation
One investigator manually (RvL) delineated bilateral carotid vessel structures of the patients and references on co-registered PET/CT slices (Inveon Research Workplace version 4.2, Siemens). She segmented the common carotid artery (extending from the aortic arch until the carotid bulb), internal carotid artery, external carotid artery and the carotid canal and put ROIs in the lumen, vessel wall and atherosclerotic plaques. The carotid bifurcation was excluded from the ROI to prevent influence of the partial volume effect. ROIs were additionally reviewed by a neuroradiologist with 12 years of experience in neuroradiology. The mean SUV was calculated per ROI and for all regions combined.
Tracer distribution
We set ROIs for blood pool, muscles and spleen to investigate whether variations in tracer distribution between COVID-19 patients and references occurred and calculated SUV mean values. As a representation for blood pool, an ellipsoid (5cm3) was drawn in the lumen of the descending aorta. Muscle activity (where uptake is expected to be low) was calculated using an ellipsoid (5cm3) in the trapezius muscle and an ellipsoid (10cm3) in the spleen.
Statistical analysis
Patient characteristics are displayed as counts and percentages and median with IQR. The mean standardized uptake value was calculated and the standard deviation. Differences in mean SUV between COVID-19 patients and references were analyzed using the Mann–Whitney U test. A paired T-test was used to calculate differences in affected versus unaffected parenchyma. Correlations between mean SUV and clinical parameters were calculated using Pearson r. Two-sided P values of less than 0.05 were considered statistically significant. Statistical analysis was performed using SPSS 25 (IBM) and Graphpad Prism 5 software (GraphPad Software).
