Abstract
609
Objectives: Total Metabolic Tumor Volume (TMTV) as assessed by FDG PET is a valuable biomarker of the disease burden. Yet, TMTV reliability is based on the accuracy of the segmentation of hypermetabolic regions, which suffers from variability due to both the differences in image quality between centers and the absence of ideal segmentation methods in PET. We thus compared the variability of TMTV estimates across 9 nuclear medicine departments (NMD) using 3 segmentation methods and 3 reconstruction schemes.
Methods: A NEMA IEC body phantom including 6 fillable spheres (inner diameter of 10, 13, 17, 22, 28, and 37 mm) was scanned in each of the 9 NMD, equipped with the Discovery 710 (5 NMD), Discovery 690 (1 NMD), Biograph 6 (1 NMD), Biograph 20 mCT (1 NMD), Biograph 40 mCT (1 NMD) scanners. The FDG sphere-to-background activity ratio was identical in each sphere and had a mean of 5.2±0.7 across the 9 NMD. The total phantom activity at the acquisition time varied from 20 to 27 MBq, corresponding to 2.5 to 3.5 MBq/kg as a function of the NMD practice. PET acquisitions (2 bed positions) were performed on each site, setting the acquisition time as a function of the NMD practice (from 120 to 210 s). Images were reconstructed on each site using 3 protocols: 2 agreed-on in advance (iterative reconstruction with scatter and attenuation corrections, a point spread function model and post-smoothing, accounting or not for TOF information) and the site-dependent Clinical routine Reconstruction Protocol (CRP). The two non-routine reconstruction protocols were defined to yield similar image quality on the GE and Siemens scanners. Image analysis was performed in a single core lab using a home-made TMTV measurement application allowing for automatic initial identification of the regions to be segmented, with no manual contouring needed to delineate the 6 volumes of interest including the spheres. Spheres segmentation was completed using a 2.5 SUV threshold, a threshold set to 40% of the local SUVmax, and using an adaptive threshold (Nestle et al 2005). The within- and between-NMD variability of the TMTV estimate (true value = 47.6±0.5 mL) was characterized by the TMTV coefficient of variation (COV).
Results: Overall, the mean bias in TMTV estimate was 9.9%±14.5 (8.9%±15 when using the CRP). The TMTV COV over all NMD, all reconstructions, and all segmentation methods was 13.5%. Within each center, the TMTV COV across the 3 reconstruction schemes for a given segmentation method was 1.6% on average (sd=1.2%). Within each center, the COV across the 3 segmentation methods when holding the reconstruction scheme fixed was 12.2% on average (sd=1.4%). When holding the segmentation method and the reconstruction fixed, the COV across the 9 NMDs was 9.9% on average (sd=1.9%). The smallest COV across NMD was systematically observed using the adaptive Nestle segmentation (mean COV of 7.6%±2), for which the mean TMTV bias was 10.4%±8.1 when using the CRP.
Conclusion: In this simple experimental setting, TMTV estimates were found to be more sensitive to the segmentation method than to the reconstruction scheme. For a given segmentation method, the TMTV estimate was well reproducible across NMD, even when using different reconstruction schemes. These results suggest that multicenter analysis of TMTV estimates is feasible if processing the data with the same segmentation approach. Research Support: n/a