Computational techniques for fast Monte Carlo validation of proton therapy treatment plans

Access to files

•   FULL-TEXT.PDF (pdf)

Abstract

Proton therapy is an established radiotherapy technique for the treatment of complexcancers. However, problems exist in the planning of treatments where the use of inaccuratedose modelling may lead to treatments being delivered which are not optimal. Most ofthe problems with dose modelling tools used in proton therapy treatment planning lie intheir treatment of processes such as multiple Coulomb scattering, therefore a technique thataccurately models such effects is preferable. Monte Carlo simulation alleviates many of theproblems in current dose models but, at present, well–validated full–physics Monte Carlosimulations require more time than is practical in clinical use.Using the well–known and well–validated Monte Carlo toolkit Geant4, an application—called PTMC—has been developed for the simulation of proton therapy treatment plans.Using PTMC, several techniques to improve throughput were developed and evaluated, includingchanges to the tracking algorithm in Geant4 and application of large scale parallelismusing novel computing architectures such as the Intel Xeon Phi co–processor. In order toquantify any differences in the dose–distributions simulated when applying these changes, anew dose comparison tool was also developed which is more suited than current techniquesfor use with Monte Carlo simulated dose distributions.Using an implementation of the Woodcock algorithm developed in this work, it is possibleto track protons through a water phantom up to eight times faster than using the PRESTAalgorithm present in Geant4, with negligible loss of accuracy. When applied to a patient simulation,the Woodcock algorithm increases throughput by up to thirty percent, though steplimitation was necessary to preserve simulation accuracy.Parallelism was implemented on an Intel Xeon Phi co–processor card, where PTMC wastested with up to 244 concurrent threads. Difficulties imposed by the limited RAM availablewere overcome through the modification of the Geant4 toolkit and through the use of a noveldose collation technique. Using a single Xeon Phi co–processor, it is possible to validate aproton therapy treatment plan in two hours; with two co–processors that simulation time ishalved. For the treatment plan tested, two Xeon Phi co–processors were roughly equivalentto a single 48–core AMD Opteron machine. The relative costs of Xeon Phi co-processors andtraditional machines are also been investigated; at present the Intel Xeon Phi co–processoris not cost competitive with standard hardware, costing around twice as much as an AMDmachine with comparable performance.Distributed parallelism was also implemented through the use of the Google ComputeEngine (GCE). A tool has been developed—called PYPE—which allows users to launchlarge clusters in the GCE to perform arbitrary compute–intensive work. PYPE was usedwith PTMC to perform rapid treatment plan validation in the GCE. Using a large cluster, itis possible to validate a proton therapy treatment plan in ten minutes at a cost of roughly$10; the same plan computed locally on a 24–thread Intel Xeon machine required five hours.As an example calculation using PYPE and PTMC, a robustness study is undertaken for aproton therapy treatment plan; this robustness study shows the usefulness of Monte Carlowhen computing dose distributions for robustness studies, and the utility of the PYPE toolto make numerous full physics Monte Carlo simulations quickly. Using the tools developedin this work, a complete treatment plan robustness study can be performed in around 26hours for a cost of less than $500, while using full–physics Monte Carlo for dose distributioncalculations.

Keyword(s)

Computational Techniques; Monte Carlo; Proton Therapy

Bibliographic metadata

Institutional metadata

Record metadata