Routine and fast verification of proton therapy...
Transcript of Routine and fast verification of proton therapy...
©2014 MFMER | slide-1
Routine and fast verification of proton therapy treatment plans with a GPU-based Monte Carlo simulationH. Wan Chan Tseung1
NCCAAPM 201529 October, Middleton [email protected]
©2014 MFMER | slide-2
Motivation• Mayo Clinic Rochester recently opened a spot
scanning proton therapy center• A second one to open in Arizona mid 2016• ~2500 patients to be treated per year• Treatment planning with Eclipse (Varian Medical
Systems, Inc)• The dosimetric accuracy of the plans needs to
be verified
©2014 MFMER | slide-3
Goals
• Gold standard for dose calculations in proton therapy is Monte Carlo (MC) simulation (e.g. FLUKA, MCNP, Geant4)
• Geant4 MC calculations of a typical proton treatment plan take several hours on a large CPU cluster
• Aims:
(1) to perform MC calculations of proton plans with <2% statistical error on the prescribed dose in around 1 minute, with comparable accuracy to Geant4
(2) to use this fast MC as an accurate, inexpensive and high throughput dose verification system
©2014 MFMER | slide-4
Development hardware
$500 x2
Gaming PC with two NVIDIA GTX680 cardsDevelopment language: CUDA C
©2014 MFMER | slide-5
Particle phase space
GPU kernel
Energy loss, multiple scattering, nuclear elastic scattering
GPU kernel
Bertini cascadeGPU kernel
Nuclear evaporationGPU kernelHost
CPU
CT image
Simulation ComponentsGeant4
(TOPAS) PBS nozzle
simulation
NB: kernel = piece of code that runs on the GPU but can be called from the CPU
©2014 MFMER | slide-6
GPU kernels for simulating non-elastic proton-nucleus interactions
intranuclearcascade
particle evaporation
• First-ever detailed simulation of nuclear interactions on a GPU
• Two kernels to handle Bertini cascade and subsequent evaporation stage
• Ability to handle proton collisions with any therapeutically relevant nucleus
• Predictions of secondary proton and neutron properties following nuclear collisions in good agreement with Geant4
More details:H. Wan Chan Tseung & C. Beltran, Computer Physics Communications, 185 (2014) 2029
©2014 MFMER | slide-7
GPU kernels for simulating non-elastic proton-nucleus interactions : results
70 MeV protons on
Carbon
70 MeV protons on
Calcium
200 MeV protons on
Carbon
200 MeV protons on
CalciumOur GPU MC 0.97 s 1.5 s 1.14 s 1.89 s
Geant4 Binary 173.69 s 317.53 s 202.28 s 461.54 s
Time taken to compute 260k nuclear interactions:
×200 speedup
©2014 MFMER | slide-8
Benchmarking our Monte CarloWe performed extensive comparisons with Geant4. Example:
230 MeV pencil beam of protons in water 200 MeV pencil beam of protons in titanium
3-D gamma pass rate (2%-2mm) : 100%
More details:H. Wan Chan Tseung, J. Ma & C. Beltran, Medical Physics, 42 (2015) 2967
©2014 MFMER | slide-9
Our GPU MC Geant4/TOPAS
Geant4/TOPASOur GPU MC
Our GPU MC Our GPU MC Geant4/TOPAS Geant4/TOPAS
60 million proton histories, 1.5% statistical error on doseGeant4/TOPAS on a CPU cluster: 650 CPU-hoursOur GPU MC: 3.5 minutes on one NVIDIA GTX680 card
Complex Head and Neck plan 3-D gamma (2%-2mm) pass rate: 98.2%
Comparisons of DVHs for various structures
©2014 MFMER | slide-10
Very fast dose-averaged LET calculation on GPU
• Our MC does simultaneous LET and physical dose calculation• Our LET calculation takes into account secondary proton contributions• Exciting prospects for MC-based RBE-weighted treatment planning
TOPAS Our GPU MCLETd LETd
TOPAS Our GPU MCdose dose
6x107 proton histories 6x107 proton
histories3 mins using 2 GTX680 cards
©2014 MFMER | slide-11
Web interface for GPU-based dose 2nd check
©2014 MFMER | slide-12
Current status• All patient plans are ran through the dose
second check • So far, ~90 patients in total• Plans modified in certain cases based on Monte
Carlo results• LET calculations allow biological dose
distribution estimates. Plans were sometimes modified based on biological dose distributions.
©2014 MFMER | slide-13
Example: Biological dose hot spot in brain stem
©2014 MFMER | slide-14
Summary
• We developed a very fast and accurate GPU-based MC for proton therapy, with detailed modeling of nuclear interactions
• This allows us to calculate dose with more confidence, especially in cases where significant hardware is present
• Compared to a 100-node CPU cluster, proton treatment plans are simulated ~100 times faster using one single graphics gaming card
• A dosimetric verification system was deployed at Mayo Clinic, using a dual GPU server
• We obtain routine, near real-time Monte Carlo feedback on pencil-beam plan accuracy
• We are exploring other exciting applications: MC-based IMPT planning, biological planning, fast neutron dose calculation, extension to heavy ions…
©2014 MFMER | slide-15
Beyond plan verification:other applications of our GPU-based MC
©2014 MFMER | slide-16
• Steps:• Determine list of spots contained in target• Pre-calculate dose influence map for these spots with GPU MC• Find spot weights with gradient-based optimization, also on GPU
• 30 minutes net calc. time on 24 GPU system: clinically applicable
• Robust IMPT optimization currently under investigation
MC-based IMPT treatment planning
GPU MC-based 3-field H&N plan
J. Ma, C. Beltran, H. Wan Chan Tseung and M. Herman, Med. Phys. 41 (2014) 121707
©2014 MFMER | slide-17
MC-based biological planning for proton therapy
• Phase 1 clinical trial on proton biological dose escalation starting end 2015
©2014 MFMER | slide-18
AcknowledgementsThis work was funded in part by Varian Medical Systems, Inc.
Thank you for your attention