|
New $2.6 Million Study Uses Video Cards to Bring Effective, Inexpensive Supercomputing to Hospitals for Safer CT Scans
Engineering and Computer Science Researchers at
Rensselaer Polytechnic Institute To Partner with GE Global
Research, Massachusetts General Hospital, and Los Alamos
National Lab
Video gamers are generally the biggest consumers of computer
graphics cards, using the devices to boost the speed and
resolution of their digital quests to fend off invading
extraterrestrials or outwit hostile dragons. But researchers at
Rensselaer Polytechnic
Institute seek to harness the power of these computer
graphics cards to solve one of the world’s most pressing health
care technology challenges: radiation exposure from X-ray and
CT imaging scans.
Rensselaer nuclear engineering expert X. George Xu is
leading an interdisciplinary team of academic, medical, and
industrial researchers to develop new techniques for quickly
calculating the radiation dose a patient will receive from a CT
scan. Funded by a $2.6 million grant from the National Institute of
Biomedical Imaging and Bioengineering (NIBIB), the research
team aims to use NVIDIA video cards and leading-edge parallel
processing techniques to help reduce radiation dose
calculations from 10 hours to less than 60 seconds. The team
also seeks to test this technology at Massachusetts General
Hospital using General Electric’s LightSpeed CT scanners.
“With this new study, we hope to bring massively parallel
computing power—currently available only to national
laboratories and major research universities such as
Rensselaer—to busy and resource-limited hospitals,” said Xu,
professor in the Department of Mechanical,
Aerospace, and Nuclear Engineering (MANE) and the Department of Biomedical
Engineering at Rensselaer, who heads the university’s Nuclear
Engineering Program. “There is a high level of interest at
the national level to quantify and reduce the amount of
ionizing radiation involved in medical imaging. Our parallel
computing method has the potential to be used in everyday
clinical procedures, which would dramatically decrease the
amount of radiation we receive from CT scans.”
Three Rensselaer faculty members are partnering with Xu on
this study: Wei Ji, assistant
professor in MANE; Christopher
Carothers, professor in the Department of Computer
Science; and Mark Shephard, the
Samuel A. Johnson ’37 and Elisabeth C. Johnson Professor of
Engineering and director of the university’s Scientific Computation
Research Center. The Rensselaer team will collaborate with
radiologist Mannudeep Kalra and medical physicist Bob Liu at Massachusetts General
Hospital in Boston, and Paul Fitzgerald from GE Global Research.
Additionally, Forrest Brown from Los Alamos National Lab will
serve as a consultant to the project.
As part of the study, the research team will perform
calculations using the Rensselaer supercomputing center, the Computational Center for
Nanotechnology Innovations (CCNI).
Medical imaging scans can provide invaluable information to
physicians, and the clinical benefits of a single X-ray or CT
scan almost always outweigh the risks of radiation exposure
from the procedure. But this radiation exposure risk
accumulates over a patient’s lifetime, and is a growing concern
given the increased frequency with which CT scans are
prescribed and performed.
A 2009 report by the National Council on Radiation
Protection and Measurements (NCRP), of which Xu is a member,
details how the U.S. population is now exposed to seven times
more radiation every year from medical imaging exams than it
was in 1980. While CT scans only account for 10 percent of
diagnostic radiological exams, the procedure contributes
disproportionately—about 67 percent—to the national collective
medical radiation exposure.
To help mitigate this risk, several national and
international bodies have called for the establishment of a
centralized “dose registry” system. Such a system would track
over time the number of CT scans a patient undergoes, and the
radiation exposure resulting from those procedures. Additional
efforts by the radiology community call for new measures to
avoid unjustified CT scans and to greatly reduce the radiation
exposure for pediatric and pregnant patients. However, current
software packages for determining and for tracking CT doses are
insufficient for such a critical task, Xu said.
To help solve this problem, Xu has spent nearly a decade
developing software that uses highly realistic 3-D virtual
reality models—called “computational phantoms”—to calculate the
exact amount of radiation a specific organ of the patient will
receive from a CT scan. Running on a standard desktop computer,
however, the software currently takes about 10 hours to perform
the Monte Carlo calculation and produce a result—far too long
to be practical in a clinical setting. Monte Carlo simulation,
originally developed in the 1940s as an outgrowth of nuclear
weapons research, is the only computational method that can
provide accurate results for non-uniform 3-D subjects like the
human body, Xu said.
In the new $2.6 million study funded by NIBIB, Xu and the
research team will design and test new Monte Carlo simulation
software to be run on the graphic processing units (GPUs) found
in computer graphics cards, instead of running solely on the
central processing units (CPUs) of a desktop computer. They
have to build the software from scratch, as no existing
radiation dose software is compatible with extremely fast
processors. GPUs are based on “stream processing” programming,
which enables efficient and effective parallel processing.
Connecting a small number of these video cards presents an
inexpensive option for users in hospitals to tackle this “Big
Data” challenge and perform massively parallel computation, Xu
said. The research team has published preliminary results
showing a single $2,000 GPU card can perform as fast as a
1,000-CPU cluster using an existing Monte Carlo code.
“The high-performance computing community is exploring the
role of GPUs in massively parallel supercomputer systems,”
Carothers said. “From a computer science perspective, in this
project we want to understand the fundamental interplay between
the algorithms used in the CT dose software and the underlying
hardware architecture of existing and forthcoming GPU
processors from NVIDIA and ‘cluster on a chip’ designs from
Intel and others. We plan to benchmark these architectures with
the CT software application against our IBM Blue Gene/Q
supercomputer system at CCNI.”
After developing and validating the software, the research
team will integrate it with GE LightSpeed CT scanner models and
a library of “phantoms” to estimate patient radiation doses in
less than a minute. Finally, to demonstrate and evaluate the
technology’s clinical benefits, the research team will perform
a series of calculations for typical diagnostic CT scanning
protocols of the head, chest, and abdomen at Massachusetts
General Hospital. Xu said he is confident this research will
one day lead to an ability to perform “patient-specific” dose
calculations using innovative image processing tools that do
not currently exist.
Nuclear engineers rely on Monte Carlo simulation as an
essential research tool. Xu said this new unique
hardware/software technology research could open the door to
many applications in clinical imaging and radiotherapy
practices as well as in the analysis of nuclear reactor systems
and the evaluation of safety of nuclear workers.
For more information on Xu’s research at Rensselaer,
visit:
For more information on parallel scientific computing research at Rensselaer,
visit:
|
Published
September 24,
2012 |
Contact: Michael Mullaney
Phone: (518) 276-6161
E-mail: mullam@rpi.edu |
|
