Acceleration of Monte Carlo CBCT Scatter-Projection Estimates Via Importance Sampling-Based Weight Windowing
A Sampson1, J Williamson2*, R Baker3, (1) Virginia Commonwealth University, Richmond, VA, (2) Virginia Commonwealth University, Richmond, VA, (3) Los Alamos National Laboratory, Los Alamos, NMWE-G-134-5 Wednesday 4:30PM - 6:00PM Room: 134
Purpose: To increase the efficiency of Monte Carlo (MC) estimation of cone-beam computed tomography (CBCT) scatter projections by incorporating systematic importance sampling- into an application-specific code.
Methods: An already optimized in-house Monte Carlo CBCT scatter-projection simulator, PTRAN, was modified to incorporate importance sampling via the weight windowing WW technique. In summary, the weight window uses Russian roulette and particle splitting to ensure that W, (the particle weight) has a value within a specified interval about target value I(P) for a phase-space location P. The WW center, I(P), is set equal to the inverse of the relative importance, defined as the ratio between the expected contribution to the detector score from particles in P, including their progeny, and the average detector score from all particles. I(P) is computed on-the-fly and is updated periodically throughout the simulation. This technique was used to compute the scatter projections in full fan mode for a simulated head phantom and in half fan mode for a simulated body phantom using a 160x120 array of 2.5 mm square detector pixels. The efficiency of importance-sampling PTRAN relative to standard PTRAN was assessed.
Results: The average efficiency gains for the simulated head and body phantom cases were 13.9±6.4 and 17.4±10.2, respectively. On a single core, the CPU time required for an average 5% standard deviation was reduced from 60 to 5 minutes (head) and 79 to 5.0 minutes (body).
Conclusion: The on-the-fly importance sampling technique dramatically improves Monte Carlo simulation efficiency and has the potential to make patient-specific CBCT scatter projection estimation clinically feasible.