Ian C. Smith1

A portal-based system for quality assurance of radiotherapy treatment plans using Grid-enabled High Performance Computing clusters

CR Baker2, V Panettieri3, C Addison1, AE Nahum3

1 Computing Services Dept, University of Liverpool; 2 Directorate of Medical Imaging and Radiotherapy, University of Liverpool; 3 Physics Department, Clatterbridge Centre for Oncology

Outline

Introduction to radiotherapy treatment planning

University of Liverpool Grid Computing Server (GCS)

GCS tools

Command line job submission using the GCS

UL-GRID Portal

Results

Future directions

Rationale

Routine radiotherapy treatment planning is constrained by lack of sufficiently powerful computing resources

Monte Carlo (MC) based codes can provide accurate absorbed dose calculations but are computationally demanding (single simulation can take 3 weeks on a desktop machine)

Fortunately MC methods are inherently parallel – can run on HPC resources and (for some codes) HTC resources

So far looked at looked running simulations on local and centrally funded HPC clusters in a user-friendly manner

Starting to look at using Condor pools

Radiotherapy codes

Two MC codes have been investigated to date: MCNPX (beta v2.7a)

general purpose transport code, tracks nearly all particles at nearly all energies (https://mcnpx.lanl.gov/).

parallel (MPI-based) code, only runs on clusters self contained – no need for pre- and post- processing steps

PENELOPE general purpose MC code implemented as a set of FORTRAN routines coupled electron-photon transport from 50 eV to 1 GeV in arbitrary materials

and complex geometries[1]. serial implementation, will run on clusters and Condor pools needs pre- and post- processing to set up input files and combine partial

results

Starting to look at EGSnrc / BEAMnrc / DOSXYZnrc[1] Salvat F, Fernández-Varea JM, Sempau J. PENELOPE, a code system for Monte Carlo simulation of electron and photon transport. France: OECD NuclearEnergy Agency, Issy-les-Moulineaux; 2008. ISBN 9264023011. Available in pdfformat at: http://www.nea.fr.

Simulation of an electron treatment: from the treatment head to the patient

(taken from Cygler et al)

Courtesy of Prof. A. Nahum (CCO)

Grid Computing Server / UL-GRID Portal

Grid Computing Server / UL-GRID software stack

Grid Computing Server tools single sign on to resources via MyProxy, use ulg-get-proxy (proxies

automatically renewed)

job management is very similar to local batch systems such as SGE: ulg-qsub, ulg-qstat, ulg-qdel etc

support for submitting large numbers of jobs, file staging and pre- and post- processing

job submission process is the same for all compute clusters (local or external)

utility tools1 provide simple Grid based extensions to standard UNIX commands: ulg-cp, ulg-ls, ulg-rm, ulg-mkdir etc.

status commands available e.g. ulg-status, ulg-rqstat

1 based on GROWL scripts from STFC Daresbury

PENELOPE (serial code) workflows

Rereasdasdascreate random seeds for N input files using

clonEasy[1]

combine N individual phase-space files

compute individual phase-space file

create random seeds for N input files using clonEasy[1]

stage-in phase-space file (only if necessary)

compute partial treatment simulation results

combine partial treatment simulation results using clonEasy[1]

repeat for other patients

phase-space filecalculation

patient treatment simulation

PortalHPC cluster

[1] Badal A and Sempau J 2006 A package of Linux scripts for the parallelization of Monte Carlo simulations Comput.Phys. Commun. 175 440–50

GCS job description files for PENELOPE (1)## create phase space file (PSF)#job_type = remote_script

host = ulgbc2total_jobs = 16name = penelopeLPO

pre_processor = /opt1/ulgrid/apps/penelope/seed_input_filespre_processor_arguments = penmain_acc6_LPO35_.in 16

indexed_input_files = penmain_acc6_LPO35_.ininput_files = spectrum_pE_6_LPO35.geo, 6MW_2.mat

executable = /usr/local/bin/run_penmainarguments= penmain_acc6_LPO35_INDEX.in penmain_LPO35_INDEX.out

log = mylogfile

GCS job description files for PENELOPE (2)## perform patient simulation using previously calculated phase space file (PSF)#job_type = remote_scripthost = ulgbc2name = penelopetotal_jobs = 10

pre_processor=/opt1/ulgrid/apps/penelope/seed_input_filespre_processor_arguments=penmain.in 10

staged_input_files=PSF_test.psfinput_remote_stage_dir=staging

input_files = water_phantom.geo,water.matindexed_input_files = penmain.in

executable = /usr/local/bin/run_penmainarguments= penmainINDEX.in penmainINDEX.out ics

log = penelope.log

Condor job files for PENELOPE # job description filegrid_resource = gt2 ulgbc2.liv.ac.uk/jobmanager-condorg universe = gridexecutable = /usr/local/bin/run_penmain arguments = penmain_acc6_LPO35_$(PROCESS).in penmain_LPO35_$(PROCESS).out ics_test+ulg_job_name = penelopeLPOlog = logtransfer_input_files = spectrum_pE_6_LPO35.geo, 6MW_2.mat, penmain_acc6_LPO35_$(PROCESS).intransfer_files = alwaystransfer_executable = FALSE GlobusRSL = (count=1)(job_type=remote_script) \

(input_working_directory=/condor_data/smithic/penelope/big_test/create_psf) \ (job_name=penelopeLPO)notification = neverqueue 16

# DAG fileJOB pre_process dummy1.subJOB staging penelopeLPO35.subSCRIPT PRE pre_process /opt1/ulgrid/apps/penelope/seed_input_filesPARENT pre_process CHILD staging

GCS job submission and monitoringsmithic(ulgp4)create_psf$ ulg-qsub penelopeLPO35 Grid job submitted successfully, Job ID is 125042

smithic(ulgp4)create_psf$ ulg-qstat

Job ID Job Name Owner State Cores Host------ -------- ----- ----- ----- ----

125015.0 penelopeLPO smithic pr 1 ulgbc2.liv.ac.uk125034.0 penelope vpanetti r 1 ulgbc2.liv.ac.uk125035.0 penelope vpanetti w 1 ulgbc2.liv.ac.uk125038.0 penelope smithic si 1 ulgbc2.liv.ac.uk125042.0 penelopeLPO smithic qw 1 ulgbc2.liv.ac.uk125043.0 mcnpx3 colinb r 64 ulgbc4.liv.ac.uk125044.0 mcnpx3 colinb r 64 lancs2.nw-grid.ac.uk125044.0 gamess_test bonarlaw r 32 ngs.rl.ac.uk

Lung treatment simulated with PENELOPE and penVOX 07

7 fields

PSF calculation 1.5 days (14 cores) approximately 1.5 million particles

Patient calculation 1.5 days for all 7 fields (single core) Statistical uncertainty 1% (1 sigma)

2.5cm diameter beam, full energy (~60 MeV at patient, ~3.2 cm range in water)

500 million histories0.5x0.5x5 mm voxels50keV proton cut-off

<1% statistical uncertainty in absorbed dose in high dose region (1s)

Bragg peak

Half-modulation

Proton absorbed dose in water using MCNPX

Future Directions

Provide support for BEAM[1] and DOSxyz[3] (based on the EGSnrc MC code [2])

Utilise Liverpool Windows Condor Pool for running PENELOPE jobs

Compare with other implementations e.g. RT-Grid.

References:

[1] 23D. W. Rogers, B. Faddegon, G. X. Ding, C. M. Ma, J. Wei, and T. Mackie, “BEAM: A Monte Carlo code to simulate radiotherapy treatment units,” Med. Phys. 22, 503–524 _1995_.

[2] Kawrakow and D. W. O. Rogers. The EGSnrc Code System: Monte Carlo simulation of electron and photon transport. Technical Report PIRS-701 (4th printing), National Research Council of Canada, Ottawa, Canada, 2003.

[3] Walters B, Kawrakow I and Rogers D W O 2007 DOSXYZnrc Users Manual Report PIRS 794 (Ottawa: National Research Council of Canada)