Metal Artifact Reduction of CT Scans to Improve PET/CT

Phantom Scan

The phantom, depicted in Figure 1A, consisted of 2 types of hip prostheses (cobalt steel and Ti-Al-V with ceramic head) and 2 solid steel balls (2.5- and 3.8-cm diameter) in a solution of ¹⁸F-FDG and water in a 9-L container with sponges to position the prostheses and balls. Background radiotracer concentration was approximately 3.2 kBq/mL, similar to values seen in patient scans. The phantom was scanned in a Biograph 40 mCT PET/CT scanner (Siemens Healthcare). The phantom was in the same position on the patient bed in the CT and PET measurements; however, to investigate the effect of motion on the iMAR algorithm an additional phantom scan was obtained for which a second CT scan was acquired after the phantom was translated 3 mm laterally in the region of the prosthetic femoral heads. We reconstructed 1 PET bed position using the misaligned CT, with a 2-mm voxel size (400 × 400 matrix size) and no postreconstruction filter to make the image as sharp as possible and artifacts as plainly visible as possible.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1.

(A) Phantom consisting of 2 types of hip prostheses in solution of ¹⁸F-FDG and water: a cobalt steel hip prosthesis and a Ti-Al-V hip prosthesis with ceramic head. Furthermore, phantom included 2 solid steel balls (2.5 and 3.8 cm in diameter). Phantom was scanned in mCT PET/CT scanner. Cold artifact caused by metal is present in standard PET reconstruction (B), whereas on iMAR PET reconstruction (C) these artifacts were not visible.

Patient Scans

PET/CT images of 21 patients with metal implants in the field of view of clinical interest were included in the study. The institutional review board of the Radboud University Medical Center approved this retrospective study, and the requirement to obtain informed consent was waived.

In 14 patients, ¹⁸F-FDG PET/CT scanning and in 7 patients ⁶⁸Ga-PSMA PET/CT scanning were performed. The administered ¹⁸F-FDG activity depended on the patient’s weight, resulting in an average and SD of 164 ± 132 MBq with a mean incubation time (and SD) of 61 ± 5 min. For ⁶⁸Ga-PSMA-HBED-CC, the administered activity also depended on the patient’s weight, resulting in 135 ± 26 MBq using an incubation time of 66 ± 17 min. Patients were scanned in a Biograph 40 mCT PET/CT scanner (Siemens Healthcare) with extended field of view (TrueV). This scanner has been accredited by the European Association of Nuclear Medicine Research Ltd. (12). Included in this study were patients with one or more metal implants with (a strong suggestion of) lesions near the implant, located in the area of the metal artifacts, scanned between January 2015 and October 2016, and consecutive patients with (the strong suggestion of) ⁶⁸Ga-PSMA–avid lesions near the implant, located in the area of the metal artifacts. The indication for the PET/CT scan differed per patient: patients were scanned for diagnosis of disease, staging, radiotherapy planning, or response monitoring. Patient characteristics are summarized in Table 1.

CT Acquisition and Reconstruction

The low-dose CT scan was acquired for attenuation correction and anatomic reference. The x-ray tube voltage was chosen using Care kV (Siemens Healthcare), with a reference value of 120 kV. The tube current was modulated using Care Dose4D (Siemens Healthcare), with a reference current–time–product of 50 mAs per rotation. The CT data were acquired with a 0.5 s rotation time and 16 × 1.2 mm collimation.

CT projection data were processed without and with iMAR. In the first case, we used kernels that we have used for several years in whole-body PET/CT. In the second case, we used the iMAR algorithm (Siemens AG). This method, which has not yet been fully described in the peer-reviewed scientific literature, is based on a sinogram metal in-painting approach for which the original sinogram values behind the metal are replaced. It consists of several steps. The first step is the so-called normalized MAR method described by Meyer et al. (13). In this step an initial prior image is estimated that consists of water and bone, posing a simplified representation of the patient anatomy without the metal artifacts. Artificial raw data are generated from this prior image and used to replace the original metal artifacts–affected raw data behind the metal. Because the metal artifacts are mainly long range, low-frequency artifacts (e.g., dark streaks) can be identified and the high frequency structural information around the metal can be used from the original image to restore correct anatomic information in the MAR image. This is the second step of the iMAR algorithm and is called the frequency-split technique (14). The resulting MAR-corrected image can be used again to estimate a more accurate prior image so that the first 2 steps can be performed iteratively (up to 5 loops depending on the chosen preset). In addition to this aggressive sinogram in-painting approach (which is effective for strong metal artifacts due to photon starvation), the iMAR algorithm also incorporates a method that corrects the beam-hardening effects for metal artifacts originating from smaller pieces of metal. The beam-hardening–corrected sinogram values are adaptively combined with the in-painted values depending on the metal attenuation strength. To guarantee the best results for different types of metal implants, distinct parameter configurations are offered to the user (so-called presets; neuro coils; dental fillings; spine, shoulder, hip, or extremity implants; pacemaker or thoracic coils) to facilitate the different iMAR-correction components. In Supplemental Figure 1 (supplemental materials are available at http://jnm.snmjournals.org), a flowchart of the different steps is shown.

The HU values are converted to PET attenuation coefficients (μ-maps) using the standard procedure described by Carney et al. (15). Because streak artifacts are decreased in soft tissue and bone, an improved PET μ-map is expected in regions close to metal. In regions not affected by metal artifact, CT images and PET μ-maps are quantitatively identical to the ones not based on iMAR. The effect of the iMAR algorithm on CT images is shown for 3 patients with different types of implants in Supplemental Figure 2. Two CT reconstructions were made for each patient, 1 without and 1 with the iMAR tool.

CT scans were reconstructed using a 5-mm slice thickness for the ¹⁸F-FDG images. The CT images of the ⁸⁶GA-PSMA scan were reconstructed with a 3-mm slice thickness, to conform to clinical protocol.

PET Acquisition and Reconstruction

The duration of the acquisition of the ¹⁸F-FDG PET scans depended on the protocol. In the whole-body protocol (head to feet), patients were scanned with 3 min per bed position, whereas in the standard protocol (head to hip) they were scanned with 4 min per bed position. The ⁶⁸Ga-PSMA patients were scanned with 4 min per bed position for the bed positions covering the pelvis; other bed positions were scanned for 3 min. Standard clinical reconstruction parameters were used for 2 PET images reconstructed with the standard CT and iMAR CT, using a 3-dimensional ordered-subset expectation maximization algorithm with a spatially varying point-spread function incorporating time-of-flight (TOF) information (UltraHD PET); 3 iterations and 21 subsets; and a slice width that was matched with the CT. A 3-dimensional gaussian filter kernel with a full width at half maximum of 3.0 mm was used as postreconstruction filtering, and a transaxial matrix size of 200 × 200 was used (4 × 4 mm² pixel size). A third non–attenuation-corrected PET reconstruction was made with similar reconstruction parameters, except for the 3-dimensional ordered-subset expectation maximization algorithm without spatially varying point-spread function and TOF (3 iterations and 24 subsets).

Image Analysis

For the image analysis, the 2 types of CT and 3 types of PET reconstructions were used. The PET/CT images were visually scored by 2 experienced nuclear medicine physicians and were also quantitatively analyzed. The physicians scored the standard, iMAR, and non–attenuation-corrected PET/CT images by answering several questions (Table 2).

The quantitative analyses consisted of measurements on both CT and PET images. For each patient scan, 6 circular ROIs were placed with a diameter of 15 mm. Two ROIs were drawn in a region in which dark streaks were visible on CT (underestimated HU), 2 in which light streaks were present (overestimated HU), and 2 on a soft-tissue location in which no metal artifact was present (background HU). The ROIs on CT and PET were matched for each patient. In addition to the SUV_mean of these 6 ROIs, the SUV_mean of the lesions located near the implant was also calculated. The spheric ROIs had diameters between 15 and 40 mm. The type of metal artifact differed for these lesions, and to illustrate the effect of the iMAR algorithm the absolute difference in SUV_mean is given for these measurements. The differences are displayed graphically as Bland–Altman plots.

Statistical Analysis

Not all the paired groups were normally distributed, and statistical analysis of the ROI measurements was performed with the Wilcoxon signed-rank test. The interrater agreement between the 2 physicians was not measured with the Cohen’s κ-coefficient because of the dichotomous nature of the questions; however, the fraction of agreement is given. Statistical analysis was performed using SPSS Statistics 22 (IBM). Statistical significance was defined for a P value less than 0.05.