NICS and JICS Collaborate with Medical Researchers to Improve Image Processing
Neurologist Frank Skidmore of the University of Alabama at Birmingham
Parkinson’s disease invades the human brain by stealth, affecting the brain chemical dopamine, which is important for controlling bodily movement. But the movement system is only the most obvious effect of this disease, and many brain chemicals eventually can be involved. Emotions, attention, memory and decision-making can all be influenced. The hallmark of Parkinson's, a misfolded protein, may start depositing years before symptoms begin. By the time those symptoms appear, about 60 to 70 percent of the dopamine-producing neurons have been destroyed.
More than one million people are living with the symptoms of Parkinson’s disease just in North America, writes physician Soania Mathur, who herself was diagnosed with the young onset occurrence of the disease at age 27. Those displaying the symptoms are the tip of the iceberg, however. Many older adults show the “footprints” of the disease, the misfolded proteins, and are at risk for developing the disease but do not yet have symptoms.
Diagnosing Parkinson’s requires the use of a series of questions and evaluation techniques, but, as Mathur explains, if what are known as biomarkers could be discovered, the effects would transform diagnosis and management of this condition. With the right biomarkers, perhaps we can interrupt the progression of the disorder before symptoms begin, develop better treatments, and, ultimately, discover a cure.
The Search for Biomarkers
The Parkinson’s Progression Markers Initiative (PPMI), a landmark observational clinical study sponsored by the Michael J. Fox Foundation for Parkinson’s Research, refers to biomarkers as a crucial missing link in the development of next-generation Parkinson’s treatments.
PPMI places biomarkers in two categories. A progressive biomarker, it states, is an objectively measurable characteristic that changes over time in a way that can be correlated to the progression of disease. The other is a diagnostic biomarker, which PPMI defines as an objectively measurable physical characteristic associated with the presence of the disease.
PPMI aims to identify one or more biomarkers of Parkinson’s disease progression. To that end, the study comprehensively evaluates the data of groups of patients of significant interest using advanced imaging, biologic sampling and clinical and behavioral assessments.
The study provided a set of what are called functional magnetic resonance imaging (fMRI) images to a research project established to structure the data in the images to make them useful in locating biomarkers.
Aligning the Structural Features
The research project team is composed of computational scientists from the National Institute for Computational Sciences (NICS) and the Joint Institute for Computational Sciences (JICS) at the University of Tennessee, and medical researchers from the University of Alabama at Birmingham (UAB). The team came together through the auspices of the Extended Collaborative Support Services (ECSS) program component of the National Science Foundation’s eXtreme Science and Engineering Discovery Environment (XSEDE). Neurologist Frank Skidmore of UAB is the project principal investigator, and Glenn Brook of JICS is the computational science lead.
The ECSS team is making advances in obtaining correct alignment of the structural features among the images in the data sets. This alignment is known as registration. The computational science work is enhancing and accelerating a procedure called high-fidelity warping that is designed to obtain better image registration. The procedure maps all positions in one image to that of another, providing enough structure to clearly see brain function.
A problem in attempts to find the commonalities in the brain images is something called inter-subject variability. Simply put, just as no face is the same, every person's brain is different—but in the case of a brain, the issue is in three dimensions. To make matters more complicated, each person may have their head positioned differently, or even may have a stroke, cyst or other finding that differentiates their brain from all the others. To use imaging for a biomarker, these differences have to be accounted for. High-fidelity warping is designed to overcome that obstacle.
The ECSS team’s computational scientists “looked under the hood” of the warping procedure to tweak the mathematical operations and improve the computer code involved in the mapping process. The code is called 3dQwarp, which divides a source image into patches and then incrementally lines up those patches with a template. The end result is a normalized (adjusted) image that can be included in statistical analyses to see which portions of the brain appear to be predictors for Parkinson’s disease.
Two brains, from two different individuals (far left, and far right) are matched to a common template (middle). Differences in head position and brain anatomy are matched using a high-fidelity warping procedure. Graphic design: Scott Gibson.
Greater Efficiency and New Discoveries
Because some of the images possessing inter-subject variability had insufficient image registration, the NICS/JICS computational scientists scaled up the 3dQwarp code to take advantage of the processing power of the Beacon advanced computing system (with its Intel® Xeon Phi™ coprocessors) at NICS/JICS.
Using the image data from PPMI, the ECSS team improved the original warping procedure by about 20 percent, they report. This greater efficiency, the team explains, increased the number of images useful for analysis from 66 percent to more than 98 percent of the research data sample. And they add that the execution of multiple streams of data simultaneously—known as multithreading in the computational science realm—on Beacon reduced computing time by 90 percent.
According to NICS computational scientist Junqi Yin: “With more data, and with better registration—particularly of small brainstem structures—we also have been able to discover previously unidentified regions that differentiate Parkinson’s disease from healthy individuals even early in the disease process. Future directions may implement a version of 3dQwarp that scales to a larger number of threads to assist in developing analytical turnaround times that will be more relevant for both research and clinical settings. Also, analysis of these and other MRI data sets using machine learning algorithms and including more covariates such as genetic information may improve diagnostic yield.”
This research, supported by the Extended Collaborative Support Services (ECSS) program of the National Science Foundation's eXtreme Science and Engineering Discovery Environment (XSEDE), has defined several regions in the brain not previously associated as affected by Parkinson's disease. Among those are the temporal tip (above, left) and the brain stem (above, right). Credits: Frank Skidmore, University of Alabama at Birmingham.
In addition to Skidmore, Brook and Yin, the ECSS team members are as follows. From UAB: Thomas Anthony and Jon Marstrander, Department of Electrical and Computer Engineering; and Yuliang Liu, Department of Biostatistics. From JICS: Chad Burdyshaw and Mitch Horton.
The team currently is developing publications about this project. The work is supported by NSF grant number ACI-1053575.
Scott Gibson, science writer, NICS, JICS
Article posting date: 26 April 2016
About JICS and NICS: Established by the University of Tennessee and Oak Ridge National Laboratory, the Joint Institute for Computational Sciences (JICS) is a conduit and a nexus for research collaborations and a provider of advanced computing resources. It also is an educator in cutting-edge computing focused on solving the most difficult problems in science and technology. JICS operates the National Institute for Computational Sciences (NICS), a leading academic supercomputing center and a major partner in the National Science Foundation's eXtreme Science and Engineering Discovery Environment (XSEDE).