¹⁸F-DCFPyL PET/CT in Patients with Subclinical Recurrence of Prostate Cancer: Effect of Lesion Size, Smoothing Filter, and Partial-Volume Correction on PROMISE Criteria

Patient Population

The cohort for this study included patients from 2 prospective institutional ethics review board–approved clinical trials from May 2018 to February 2019 (NCT03718260 and NCT03160794). Written informed consent was obtained from all subjects. In both trials, subjects who had completed primary therapy experienced biochemical recurrence with subclinical metastasis, which was defined as asymptomatic recurrence with negative or equivocal findings on conventional imaging (bone scintigraphy and CT). CT and bone scintigraphy were interpreted clinically before study accrual. Patients underwent ¹⁸F-DCFPyL PET/CT between May 2018 and February 2019.

¹⁸F-DCFPyL PET Imaging

¹⁸F-DCFPyL was synthesized as previously described (8). PET was performed on 1 of 2 scanners (mCT40 [Siemens Healthcare] or Discovery 610 [GE Healthcare]) approximately 120 min (mean, 115.1 min; range, 83–168 min) after intravenous injection of 335.5 MBq (range, 223–376 MBq) of ¹⁸F-DCFPyL, a second-generation low-molecular-weight PSMA PET radiotracer. The patients were asked to void the bladder and were subsequently positioned supine on the imaging couch with the arms outside the region of interest. Images were obtained from the top of the skull to the upper thighs. Iodinated oral contrast material was administered for bowel opacification; no intravenous iodinated contrast material was used. The acquisition time was 8 min per bed position for the pelvis and 2 min per bed position for all other sites. Overall, 5–9 bed positions were obtained, depending on patient height. For the mCT40, the CT parameters were 120 kV, a 5.0-mm slice width, a 4.0-mm collimation, a 0.8-s rotation time, an 8.4-mm feed per rotation, and a B30s kernel medium smoothing reconstruction. A PET emission scan using time of flight with scatter correction was obtained covering the identical transverse field of view, with a 2.6-pixel image size, a 3.27-mm slice width, and a gaussian filter of 5 mm in full width at half maximum with 31 iterations and 21 subsets. For the Discovery 610, the CT parameters were 120 kV, 150–190 mA depending on body weight, a 3.75-mm helical thickness, an 0.8-s rotation time, an 11.25-mm feed per rotation, and standard reconstruction at a 2.5-mm slice width. The PET parameters were a 3.18-pixel image size and a gaussian filter of 6.4 mm in full width at half maximum with 2 iterations and 32 subsets. The scanning parameters were the clinical standard at each site and were selected according to vendor recommendations for each scanner.

Image Analysis

All suspected tumor foci on ¹⁸F-DCFPyL PET/CT were manually contoured on the unenhanced CT images (acquired at the time of PET/CT) and, independently, on the 5-mm gaussian-smoothed PET images, using a Mirada Vision workstation (Mirada Medical). Two readers worked together in consensus, one performing the contouring and the other reviewing it. For all lesions, SUV_max and SUV_mean were recorded. A threshold of 41% of SUV_max was used to generate a contoured volume (mm³). This threshold was used because it has been previously suggested to represent the optimal volume match (9). For a reference, the SUV_max and SUV_mean of the right parotid gland, liver (as measured over the right hepatic lobe), and mediastinal blood pool were recorded. Each lesion received a miPSMA score (before the correction was performed) as per the PROMISE criteria.

Correction Methodology

For all analyses in this work, unsmoothed PET image data were used. Because the data were collated retrospectively from a cohort of patients in ongoing clinical trials, postreconstruction smoothing had already been applied to the images. The unsmoothed data were obtained by Richardson–Lucy deconvolution (10,11). We validated this method within a subset of patients having available PET raw data, which were directly reconstructed without a smoothing filter for comparison.

CT and PET mask volumes were generated from the contours using slice-by-slice selection of the pixels contained within the region contour points. To account for local alignment inconsistencies between PET and CT images, for each lesion in the CT data we found translational offsets that maximized the region-of-interest SUV_peak within the activity in the PET volume delineated by the corresponding region in the PET mask. The CT mask was then transformed to the PET space by linear interpolation onto the PET voxel coordinate grid, accounting for any necessary translational adjustments. Lesion volume and SUV data were recorded from the contoured lesions (i.e., the original dataset).

The matched anatomic mask was the basis for calculating the PVC factors. However, before the corrections factors could be calculated, the relative contrast for each node needed to be represented in the mask, since local differences between region and background values determine partial-volume spill-over effects. This step was accomplished through an empiric searching approach; every region was scaled by a large range of contrast values, each time followed by calculation of its respective PVC map. The correction maps were then applied to the PET volumes, and the resulting corrected activity distributions were summed over all affected voxels. The points of the sum data were fit to a curve, and the location on the curve that equaled the corresponding sum of the original data indicated that the total mass of activity had been preserved and hence yielded the correct scaling factor. The method outlined here describes an approach for PVC for lymph nodes in PET images, with the assumption of uniformly distributed activity within small volumes.

Once each region of interest in the mask was corrected and scaled appropriately against a unitary background, a map of the local PVC factors was generated. Partial-volume effects were first simulated in the anatomic mask by filtering the volume by a 3-dimensional smoothing kernel consistent with the intrinsic scanner response function (in this experiment, a gaussian kernel of 4 mm in full width at half maximum was used to model the point-spread response). Dividing the original anatomic volume by the smoothed data yielded a map of correction factors (CFs), which were then applied directly to the PET volume.

The effects of the PVC were evaluated for all regions in both unsmoothed and 5-mm gaussian smoothed images. The original versus corrected SUV_max and SUV_mean were compared.

Subsequently, the prospectively assigned miPSMA score was compared with the score obtained from the corrected dataset (Fig. 1). All quantitative and qualitative assessments were correlated for lesion volume. Furthermore, a CF was calculated as follows: [SUV_max with PVC correction]/[original SUV_max] for each lesion. To determine whether differences in pixel sizes from the 2 scanners have a significant impact on PVC method, the datasets were also analyzed separately for each scanner.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1.

A 75-y-old man with history of pT3a Gleason 7 (3 + 4) prostate cancer after radical prostatectomy 11 y ago, and salvage radiotherapy for biochemical recurrence 7 y ago, now with slowly rising PSA (2.1 ng/mL). Interaortocaval lymph node (left panel, arrow) had volume of 118.3 cm³ on CT. SUV_max measured on PET was 7.0 with miPSMA score of 1 (middle panel, arrow). After correction, SUV_max was 22.6 with miPSMA score of 3 (right panel, arrow).

Because, in routine clinical work, it is not practical to contour each node, the mean CF for lesions according to size groups was calculated. The size groups chosen were nodes larger than 9 mm along the short axis (considered abnormal on morphologic imaging), nodes smaller than 4 mm (within range of scanner spatial resolution), and 2 categories for nodes between 4 and 9 mm (4–7 mm and 7–9 mm). This step was done to generate a rule-of-thumb (RoT) CF according to lesion size that could be used clinically. As a validation, the adjusted dataset using the mean CF per lesion size group was compared with the corrected SUV dataset. Similarly, the miPSMA scores were adjusted using unsmoothed reference tissues and were compared with the corrected dataset following the same methodology.

Statistical Analysis

Statistical analysis was performed using Past software, version 3.26. Paired Student t testing was used to compare lesion SUV_max before and after corrections for the entire cohort and for each lesion size group. PSMA scores before and after corrections were compared for each size group using χ² testing. A P value of 0.05 or less was deemed statistically significant.