Program Information
GPU-Accelerated High Performance Computing Framework for IMPT Multi-Criterion Optimization
H Kamal Sayed*, M Herman , C Beltran ,Mayo Clinic, Rochester, MN
Presentations
SA-B-BRA|B-4 (Saturday, March 18, 2017) 10:30 AM - 12:30 PM Room: Ballroom A|B
Purpose: Provide clinically viable IMPT multicriteria optimization (MCO) based on GPU-accelerated high performance computing cluster (HPC)
Methods: We developed a novel MCO framework that generates Pareto plans for a set of DVH based objectives. We introduced a variation of the weighted metric method with an evolutionary adaptive weighting scheme that handles clinical DVH objectives. The adaptive weighting scheme improved the search efficiency in the multiobjective space. The MCO was developed on a GPU accelerated HPC cluster. The MCO performs multiple independent searches in parallel which results in a rapid mapping of the Pareto hypersurface. The implementation was developed using CUDA libraries for GPU programing, and MPI libraries for node communication which utilized Infiniband technology for fast data transfer. The MCO framework was implemented on multi-GPU cluster with 24 Nvidia-GeForce-GTX-TITAN cards, and 4 Intel-Xeon-CPU2.40GHz processors. Each node has 8 GPU cards.
Results: MCO was performed on a number of diverse clinical cases (head & neck, Pediatric orbit, prostate). All of the cases computation time was within 2-3 hours; which is considered the first clinically viable MCO system which provide a solution for direct machine parameters with no approximation or post processing requirements; i.e. any Pareto plan could provide machine parameters directly, hence DVH performance is persevered after MCO navigation. The MCO framework provided a set of alternative plans that showed improved DVH objectives in comparison to the clinically optimized and treated plans. The DVH improvements are shown in the attached figures. MCO plans shows OAR’s sparing with reduction of mean and host spot doses in the range of 50%-10% from the clinically treated plan.
Conclusion: We developed novel MCO for IMPT implemented on a GPU accelerated HPC cluster. The system provides rapid mapping of Pareto hypersurface. The MCO tradeoff database of IMPT plans are produced in a clinically viable timeframe.
Contact Email: