A Practical Energy-Based Scatter estimation for quantitative PET

Introduction: Scattered and Random events lead to bias in the PET images. Randoms are commonly estimated using the delayed window method. Single Scatter Simulation (SSS) is routinely used for scatter estimation where it is assumed that only one of photons scatters once. While SSS is satisfactory in many routine clinical scenarios, it does not model multiple scatter, and tail-fitting is typically needed to scale the estimate. These limitations become more marked with larger patients and/or limited counts. Here, we present a practical implementation for estimating the scattered events based on their energy properties, focusing on algorithm robustness and accuracy for routine use. In particular, we take advantage of the marked difference between the non-scattered and scattered events energy spectra to obtain an estimate for each line-of-response (LOR).

Methods: Our method uses the global energy spectrum to estimate the energy PDFs for scattered and non-scattered events. We estimate the scatter distribution by subtracting a single Gaussian kernel, fitted to the high-energy tail of events above 550 keV, from the global energy distribution. A moments-based method is then used to fit the scattered and non-scattered PDFs to the energies of all events in each LOR. The presence of random events requires care as their scatter distribution is non-correlated and different from the prompts’. Hence, we first estimate the delayed scattered PDF and then appropriately subtract it from the prompts before estimating the scatter PDF. The moments method is first used on the delayed data to estimate the random events, and then subsequently on the prompt-random data for scatter estimation. For an initial evaluation, phantom data were generated with GATE. The scanner model was based on a single ring of the PennPET Explorer scanner geometry. We used a 35 cm diameter x 70 cm long cylindrical phantom with a rectangular 2.5 cm polyethylene slab. Since the presence of randoms indirectly affects the scatter estimation, we tested without randoms and a random fraction (RF) of 20%.

Results: With 20% RF, the estimated scatter sinograms show good agreement with the ideal sinogram extracted from the Monte Carlo data. The reconstructed images show good uniformity in the background activity and a very low residual activity level in the phantom’s cold slab region due to positron range and parallax effects. Comparisons with ideal profiles, produced only by true-only (non-scattered) events, show that the method is quantitatively accurate. The mean value and background variability (BV) from multiple randomly placed ROIs was measured 0.70 (BV: 1.19%), 0.68 (1.81%) and 0.69 (1.94%), for the ideal, no-randoms, and 20% RF case.

Conclusions: We have demonstrated through realistic studies with a challenging phantom that our energy-based scatter estimation method produces accurate, quantitative images. Additional work evaluating the impact of energy window, count levels, and RF on the method's accuracy is ongoing. Results will be presented using both simulations and measurements performed on a single ring of the PennPET Explorer scanner. Acknowledgements: This work was supported in part by NIH grants R21-CA239177, R01-EB028764, R01-CA196528, R01-CA113941.

• Download figure
• Open in new tab
• Download powerpoint