(Thursday, 8/2/2018) 1:00 PM - 3:00 PM
Room: Karl Dean Ballroom A1
Purpose: Current radiotherapy divides prescribed dose equally among multiple fractions. However, various studies indicate that inter-fraction dose modulation improves dose-sparing of critical organs. Such non-uniform fractionation requires the use of Biologically effective dose â‚?BEDâ‚Ž to calculate the collective effect of fractional doses. However, the use of BED makes the optimization highly nonconvex and computationally expensive. Towards making this problem computationally tractable, we implement a non-uniform spatio-temporal fractionation â‚?NSTFâ‚Ž IMRT planning strategy, using a highly-parallelized, particle swarm optimization â‚?PSOâ‚Ž algorithm, on a prototype graphic processing unit â‚?GPUâ‚Ž-enabled treatment planning system â‚?TPSâ‚Ž. We demonstrate proof-of-concept of NSTF IMRT for a non-small cell lung cancer â‚?NSCLCâ‚Ž radiotherapy case.
Methods: We used an in-house PSO engine to optimize control point monitor units for all beams across all fractions. Eclipse �V13.6– Varian₎ TPS was used to generate and export dose matrices per control point. Universal survival curve was used for BED calculation. Several NSTF plans were created on an NSCLC patient dataset, varying fractions �3≤fx≤30₎, GPUs �1≤N≤8₎ and PSO particles �50≤P≤200₎ for a typical 9-beam IMRT plan; considering 166 control points/beam �1494 × fx optimization variables₎. Our hardware consisted of dual 8-core-Xeon CPUs and 4 Tesla-K80 GPUs �4992 cores and 24GB RAM/card₎.
Results: Our NSTF TPS reduced maximum BED for heart, esophagus and spinal cord by 50%, 51%, and 15%, respectively, compared to the clinical plan in a 5- fx, 12Gy/ fx IMRT plan, while maintaining target coverage. Computation time for 200 PSO particles was 8.1-11.2min/iteration using 7-1 GPUs, respectively. Employing >7 GPUs increased the processing time up to 8.6min/iteration due to CPU-GPU data transfer overload.
Conclusion: We developed and evaluated a GPU-enabled research Eclipse TPS for BED-based NSTF planning. The prototype demonstrated acceptable processing time â‚?2 hoursâ‚Ž despite the computational burden being over an order of magnitude larger than a clinical IMRT plan.
Funding Support, Disclosures, and Conflict of Interest: This study was partly supported by an NIH grant (R01CA169102) and a grant from Varian Medical Systems.
Keywords
Inverse Planning, Optimization
Taxonomy
TH- Radiobiology(RBio)/Biology(Bio): RBio- Photons
Contact Email
