Paralllisation dun planning de traitement en radiothrapie sur grille

1
Parallélisation dun planning de traitement en
radiothérapie sur grille

• Crédit Lydia MAIGNE
• Laboratoire de Physique corpusculaire

2
La radiothérapie est couramment utilisée pour
traiter le cancer
3
Un meilleur traitement requiert un meilleur
planning

• Aujourdhui calcul analytique de la
  distribution de dose dans la tumeur
• Pour les nouveaux traitements par modulation
  dintensité, les calculs analytiques ont une
  précision de lordre de 10 à 20 près des
  hétérogénéités
• Meilleure alternative calculs Monte Carlo (MC)
• Durée inacceptable
• Limpact des grilles réduire le temps de calcul
  MC à un niveau acceptable par un médecin

4
Calcul dun traitement de radiothérapie Un
médecin traite un patient atteint dune tumeur au
cerveau
5
1) Obtenir les images médicales de la tumeur

• Images scanner du cerveau du patient au format
  DICOM

Scanner slices DICOM format

• Format des coupes
• 512 X 512 X 1 pixels
• Taille dun voxel de limage
• 0,625 X 0,625 X 1,25 mm

6
2) Concatenate these slices in order to obtain a
3D matrix
Scanner slices DICOM format
Concatenation
7
3) Mandatory step anonymisation of the medical
data
Binary image of the scanner slices
Scanner slices DICOM format
Concatenation
Size of the matrix
Anonymisation
Size of the pixels
Image text file
Number of slices
Binary file Image.raw Size 19M
8
4) Register and replicate the binary image on
SEs
Site 1
Scanner slices DICOM format
Site 2
User interface
Concatenation
Site 3
Anonymisation
Image text file
Site 4
Binary file Image.raw Size 19M
9
5) Register the lfn of an image
Site1
Register lfn in the database
Query a lfn to the database
Scanner slices DICOM format
Site2
User interface
Concatenation
Site3
Anonymisation
Database
Image text file
Site4
Binary file Image.raw Size 19M
10
6) Split the simulations

• A GATE simulation generating a lot of particles
  in matter could take a very long time to run on a
  single processor
• So, the big simulation generating 10M of
  particles is divided into little ones, for
  example
• 10 simulations generating 1M of particles
• 20 simulations generating 500000 particles
• 50 simulations generating 200000 particles
• All the other files needed to launch Monte Carlo
  simulations are automatically created with the
  program.

jdl files jobXXX.jdl
Status files statusXXX.rndm
Script files scriptXXX.csh
Macro files macroXXX.mac
Required files
11
7) Web portal GENIUS (Grid Enabled web
eNvironment for site Independent User job
Submission)
Complex Command Line Interfaces to have access
to the Grid
Specific description of the jobs to learn (jdl
files)

• Very discouraging for potential users

Development of the GENIUS portal by INFN and
NICE company (Italy)
Development of a GATE area on the GENIUS portal
12
Jobs management and Grid interface
Registration and management of medical images

• Registration and replication anonym medical
  images
• One logical file name corresponding to multiple
  physical images
• Automatic adding and suppression of images

Jobs submissions and management
Internet connexion

• User secured authentication
• Splitting of simulations
• Automatic files creation submitted to the grid
• Jobs submission
• Jobs management
• Automatic data retrieving (spectrum, isodoses)

Working station Starting of the installation at
Centre Jean Perrin
13
A typical jdl file (with voxelized tumour image)
14
A typical jdl file (with Root analysis)
15
A typical script file (with voxelized tumour
image)
16
A typical script file (with Root analysis)
17
7) Submission on the Grid
Site1
Site2
Scanner slices DICOM format
User interface
Site3
Concatenation
Anonymisation
Database
Image text file
Site4
Binary file Image.raw Size 19M
18
Handling random numbers generation in a Grid
environment

• The random numbers generator (RNG) in MC
  simulations
• Based on deterministic algorithm
• Characteristics (example for HEPJamesRandom the
  GATE RNG)
• Very long period RNG 2144
• Creation of 900 million sub-sequences non
  overlapping with a length of 1030
• Pre-generation of random numbers
• The Sequence Splitting Method
• Until now, 200 status files generated with a
  length of 3.1010

Status 1
Status 2
Status 3
status000.rndm
status001.rndm
status002.rndm
Each status file is sent on the grid with a GATE
simulation
19
8) Analysis of output root files

• Typical dosimetry
• Merging of all the root files
• Computation of the root data

Brain_radioth000.root 20 MB
Brain_radioth001.root 20 MB
Brain_radioth002.root 20 MB
Brain_radioth003.root 20 MB
Brain_radioth004.root 20 MB
Brain_radioth005.root 20 MB
Brain_radioth006.root 20 MB
Brain_radioth007.root 20 MB
Brain_radioth008.root 20 MB
Brain_radioth009.root 20 MB
transversal view
Centre Jean Perrin Clermont-Ferrand
20
Conclusion and future prospects

• The parallelization of GATE on the DataGrid
  testbed has shown significant gain in computing
  time (factor 10)
• Improvements needed for clinical routine
• Need for guaranteed response time
• Need for service level agreement with network and
  resource providers

Ocular brachytherapy simulation with 10M of events
Parallel submission of 100 jobs
Comparison of computing time