Effects of image noise and reconstruction parameters on tumor dosimetry using 90Y PET/CT

Objectives 90Y PET is now being used to assess absorbed dose distribution in tumor and liver following 90Y-microsphere therapies. However, Y90 PET images are traditionally noisy due to the low yield of positron (32 ppm). Clinically, 90Y PET images are acquired for 15-20 min/bed to compensate for the noise. Nonetheless, accuracy and potential biases in 90Y dose estimates due to image noise and convergence have not been fully investigated. In this work we acquire “noise-free” 90Y PET images to optimize image reconstruction parameters for quantitative 90Y-PET/CT study acquired under various clinical imaging conditions: sphere-to-background ratios (SBR), sphere sizes, and acquisition duration.

Methods We used a modified IEC phantom containing 2 sphere sets of diameters 37, 17, 13mm with activity concentrations (AC) of 142 uCi/mL and 47 uCi/mL. Two different AC for background 5 uCi/mL and 11 uCi/mL were used to provide four separate SBRs of 28, 9, 13, and 4. PET data were acquired in list-mode using GE D690 PET/CT scanner for 300 min in 1-bed position. PET data was replayed to 60, 45, 30, and 15 min with 5 different realizations each. The PET/CT images were reconstructed using 3D-OSEM with PSF+TOF corrections. Total iterations varied from 12-192 to investigate convergence. The voxel size was 2.6 mm. The post-reconstruction Gaussian filters were varied from 2.6, 5.2, and 7.8 mm to investigate filtration. Absorbed dose contributions from Y90 decay are dominated by beta particles whose mean path length in water and soft-tissue is 3-5 mm. Therefore, reconstructed 90Y PET voxel-level AC in uCi/mL is proportional to voxel-level absorbed dose in Gy. Recovery coefficients (RC) of sphere AC was calculated by computing the mean AC in sphere VOIs normalized to the true sphere AC under a variety of conditions for analysis. RC is directly proportional to the bias in absorbed dose quantification. The sphere VOIs contoured using CT matched the actual size of sphere.

Results AC of all spheres converged at 48 iterations independent of size and SBR. However, the converged RC depended on sphere size, SBR, and filtrations. Mean RC for spheres was insensitive to acquisition duration (variations<5%) however image noise was strongly modulated by duration. At duration of 60 and 30min, the standard deviations in 17mm sphere images with 2.6mm filtration were 45% and 65%, respectively; this implies that while mean tumor dose is robust, dose volume histograms calculated from Y90 PET is strongly dependent on acquisition duration. In addition, shorter duration introduced non-uniformity in the images. The RC was also found to depend of the SBR when all other conditions were fixed. RC for 37 mm and 17 mm sphere varied between 75-70% and 65-50% depending on the SBR. Since absorbed dose metrics to entire tumor is desired in 90Y therapies tumor VOIs cover the entire tumor. However, sphere VOIs size strongly affected the sphere AC; the RC for 17 mm sphere varied from 50-85% depending on the VOI size. Convergence rate did not change with post-reconstruction filtration; but filtration reduced the RCs and image non-uniformity. At 30 min duration and 7.8 mm filtration, the RCs of the 17 mm sphere images were 49±2% and 54±3% for SBR of 13 and 4, respectively.

Conclusions Under tested imaging conditions, the sphere images have all converged at 48 total iterations for OSEM with PSF+TOF correction. However, the accuracy of 90Y activity quantification in terms of biases in mean tumor dose and dose volume histograms varies substantially with the imaging condition. Caution is warranted in interpreting tumor doses based on 90Y PET. Research Support: NIH/NCI R01 CA138986.