Non-cluster TomoTherapy Treatment Planning System

Abstract

Besides the Planning Station (PS) computer, the current TomoTherapy Treatment Planning System (TPS) includes a separate computer cluster with 7-14 nodes for optimization and dose calculation. We developed a novel Planning-Station-only system for TomoTherapy treatment planning. Compared with the cluster-TPS, the new solution will significantly reduce the cost and improve in both plan quality and planning throughput. The current TomoTherapy plan optimization requires pre-calculation and storage of large amount of beamlets. A computer cluster is used for both computation and data storage power to accommodate the very large scale (VSL) optimization problem. In this work, we developed a new direct-machine-parameter-optimization (DMPO) scheme that features a non-voxel and broad-beam (NVBB) representation and does not rely on beamlets. Low-memory, full computation and data parallelization nature of this scheme facilitate its efficient implementation on the graphic processing unit (GPU). We incorporated the NVBB approach into TomoTherapy TPS. The dose calculation and optimization engine runs on the same PS computer. The graphic card of the PS computer is replaced by a Navida GeForce GTX295 (GPU) card. Extensive verification and validation tests were performed in house and via third parties. Benchmarks on dose accuracy, plan quality and throughput were compared with the commercial TomoTherapy TPS with the 14-blade DC3 cluster. Clinical case studies include the prostate, lung, breast, H&N and TBM for both TomoHelical and TomoDirect planning. Compared with the current cluster-TPS, the new GPU-TPS reduced the pre-processing time from 10-200 minutes to 10 seconds as no beamlet pre-calculation is needed any longer. The iteration time was reduced to 25-90%. Drivability and plan qualities were indistinguishable for most cases and less dose artifacts were observable for a few cases via GPU-TPS. For the same delivery plan in full dose and final dose calculation, the GPU-TPS had speedup of about 8-16 times compared with the cluster-TPS, while the dose differences between cluster-TPS and GPU-TPS were within 1%, 1 mm for all test cases. The DMPO nature of the NVBB framework eliminates the needs of beamlets and leads to better plan quality. Extensive planning and benchmark studies validate the GPU-TPS. The non-cluster solution results in significant savings on the hardware and service cost. Compared with the cluster-TPS, planning time was reduced in many folds with the new GPU-TPS. With this new technique, VSL TomoTherapy treatment planning can even be accomplished via a single laptop.