Facilitating Better Breast Cancer Diagnostics
XSEDE Supercomputers Help Researchers Investigate Ultrasound Behavior in Human Breast Tissue
Breast cancer. Not only is it one of the most common cancers to afflict American women, it also is a stealthy disease that often presents no symptoms in the early stages. Its insidious nature makes regular screening and accurate tests essential.
The first line of detection is an x-ray picture of the breast, known as a mammogram. A screening mammogram usually takes two x-ray images from different angles and is the typical yearly mammogram women over age 40 get. If an abnormal result from a screening mammogram occurs or if symptoms are present, a diagnostic mammogram is used. This mammogram takes more images of the suspicious area and can capture specially magnified or shaped images to make a small area of abnormal breast tissue easier to evaluate.
Women who have a high risk of developing breast cancer or areas of concern in the breast, may be given an MRI (magnetic resonance imaging) scan. An MRI uses a magnetic field and pulses of radio waves to create pictures of breast tissue. In some women diagnosed with breast cancer, an MRI scan is done to determine the size of the cancer and see whether cancer exists elsewhere in the breast.
Another scan usually done to get a closer look at a potentially problematic area in the breast is an ultrasound. During an ultrasound, a transducer emits sound waves and picks up echoes as they bounce off body tissue. Ultrasound helps differentiate between cysts, which are fluid-filled sacs, and solid masses.
Sometimes ultrasound can even help distinguish between benign and cancerous tumors as well. In addition, ultrasound can be used to look for enlarged lymph nodes under the arms, typically the first areas in which breast cancer spreads. Currently, the use of ultrasound in lieu of mammograms for breast cancer screening is not recommended.
Current Limitations and New Possibilities
The resolution of an ultrasound is limited by acoustic scattering in the breast tissue. Accurate and efficient modeling of ultrasound propagation through realistic tissue models is important in the advancement of multiple aspects of clinical ultrasound imaging.
To address the challenge, a group of researchers from the University of Rochester applied high-performance computing resources from the Extreme Science and Engineering Discovery Environment (XSEDE) (project number TG-DMS110016) in an investigation aimed at quantifying resolution limits in the presence of acoustic beam deviation in tissue, referred to as aberration, and with the use of aberrated correction. The researchers chose to explore the viability of two different methods of modeling ultrasound scattering.
For this project, they developed models of breast tissue by segmenting magnetic resonance images of ex vivo (Latin for “out of the living”) tissue specimens. Once it was confirmed that these models mimicked in vivo (Latin for “within the living”) breast tissue, they were mapped via ultrasound, recording variations in speed and sound absorption.
The researchers mapped breast tissue samples using the fast multipole method (FMM) and the k-space method to test the validity of both methods. The group’s Andrew Hesford explains that “because the two methods are fundamentally different, the similarity of the results gives us confidence that the results are correct and representative of ultrasound behavior in typical human breast.”
Techniques for Validating Ultrasound Imaging
FMM is a mathematical technique used to produce data for scattering problems of extremely large size. In this project, FMM was used to obtain the monochromatic solutions. The k-space method computes scattering from the magnetic resonance image using difference equations and frequently refers to the temporary image space. These are large-scale, wave-propagation and scattering problems that require supercomputing systems to efficiently run. To accomplish their task, Hesford and his collaborators used the Kraken supercomputer (now decommissioned) at the National Institute for Computational Sciences and the Longhorn machine at the Texas Advanced Computing Center (TACC).
According to Hesford, typical simulations of ultrasound propagation through their breast models required about 256 Kraken nodes with 12 CPUs (central processing units) each, for a total of 3072 CPUs. “Each run simulated propagation of a single frequency,” Hesford says. “As the frequency increased, the runtime increased; at 3 MHz, where most of our work was done, the 3072 CPUs finished their tasks in about 90 minutes. At 5 MHz, the tasks took 5 hours.” Due to the variety of different configurations and the varying ultrasound frequencies, more than 1 million CPU hours were used in their project.
Computational Power Needed
Hesford shared the vital importance of using supercomputers such as Kraken and Longhorn for the team’s research. In terms of solution times, he says Kraken hours equate to more than a century on a single CPU machine and to at least a few decades on a modern multi-core systems.
“There are also memory constraints that make these simple projections either completely impossible or immensely more impractical,” Hesford says. “Our group at the University of Rochester has developed improved algorithms that can compute, in minutes, on a single computer with a high-performance graphics processing unit (GPU), a simplified ultrasound propagation through our 2.5 billion unknown breast models. The solution is accurate [under the right conditions] to within 10% of the FMM solution that would require tens of thousands of CPU hours on Kraken. While 10-percent error can be too much for certain applications, for others (such as imaging), 10-percent error is more than a reasonable tradeoff, given the massive reduction in computational complexity. What matters is understanding the limits of the methods and applying them judiciously.”
Hesford says that ideally their work with this project will lead to “perfecting simplified methods that can provide similar results with greatly reduced computational demands.” Currently the computational price and the resources required eliminate this modeling from being directly useful in medical imaging applications, he explains. But he adds that these results can serve as reference standards that will enable the researchers to use less computationally intensive models of acoustic wave behavior for imaging.
According to Hesford, this research could lead to better diagnostic techniques and eventually ultrasound as a viable diagnosis alternative because “the reference results produced now allow us to evaluate a variety of simplified physical models in context of very complicated acoustic behavior inside human breast tissue.”
A more detailed look at the researchers’ study can be found in the Journal of the Acoustical Society of America under the title “Comparison of temporal and spectral methods using acoustically large breast models derived from magnetic resonance images.”
Current U.S. breast cancer statistics indicate that about one in eight women will develop invasive breast cancer over the course of their lifetime and about one in 1,000 men. Aside from lung cancer, breast cancer death rates are higher for U.S. women than any other cancer. About 85 percent of breast cancers occur in women who have no family history of breast cancer.
Since early detection is crucial, women should begin performing breast self-exam while in their 20s to know how their breasts normally look and feel so they can promptly report any changes to their doctor. A clinical breast exam should occur every three years for women in their 20s and 30s. Women in their 40s and older should have a clinical breast exam and a mammogram every year.
This research was funded in part by grants R01 EB009692 and R01 EB010069 from the National Institute of Biomedical Imaging and Bioengineering in the National Institutes of Health. Calculations associated with the work were made possible by project TG-DMS110016 from the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number OCI-1053575.
Jennifer Bailey, science writer, NICS, JICS
Article posting date: 8 January 2015
Last access date: 5 February 2015
About JICS and NICS: The Joint Institute for Computational Sciences (JICS) was established by the University of Tennessee and Oak Ridge National Laboratory (ORNL) to advance scientific discovery and state-of-the-art engineering, and to further knowledge of computational modeling and simulation. JICS realizes its vision by taking full advantage of petascale-and-beyond computers housed at ORNL and by educating a new generation of scientists and engineers well versed in the application of computational modeling and simulation for solving the most challenging scientific and engineering problems. JICS runs the National Institute for Computational Sciences (NICS), which had the distinction of deploying and managing the Kraken supercomputer. NICS is a leading academic supercomputing center and a major partner in the National Science Foundation’s eXtreme Science and Engineering Discovery Environment, known as XSEDE. In November 2012, JICS sited the Beacon system, which set a record for power efficiency and captured the number one position on the Green500 list of the most energy-efficient computers.