New imaging technology developed using Rice cluster could make images from CT scans better, faster
December 18, 2015
Sharper images to aid diagnoses from medical scans that also can be generated faster, making them less expensive and exposing patients to lower doses of radiation, could be one benefit of technology being developed by campus researchers using Purdue’s latest community cluster supercomputer Rice.
Xiao Wang, an electrical and computer engineering doctoral student in the lab of Professor Charles Bouman, is focusing on improvements in model-based image reconstruction (MBIR) used in conjunction with computed tomography (CT) scanning, referred to commonly as a “cat” scan.
The traditional method of imaging from CT scans generates images that include a lot of noise, artifacts from the system and other extraneous information, and make the picture grainy. This makes it harder for a physician to discern, for example, a tumor amid the visual clutter.
Model-based image reconstruction (MBIR) applies computer models to CT data designed to filter out the noise and sharpen the details, as sharp as the more expensive imaging method Magnetic Resonance Imaging (MRI).
“Usually MBIR has much better image quality, but the main problem is it takes lots of time to finish the reconstruction work, which makes the cost higher,” Wang says
Wang and his colleagues in Bouman’s lab are looking at how to make the computations run as fast as possible, lowering the cost and saving patients money by offering an alternative to MRI.
MBIR has applications beyond medical imaging. The Department of Homeland Security funds the Purdue research, in part, because of its application in CT scanning of baggage and other cargo at airports. Commercially, it is being used to develop cameras that can actually focus a picture after it is taken.
A paper by Wang, Bouman and Purdue Professor Samuel Midkiff, presented at the SC15 international supercomputing conference in November, outlined a method for generating a 3-D image from a CT scan using MBIR that is significantly faster — up to 187 times faster in some cases. It was selected as one of the top papers at the conference.
“This is a very important application domain and 187 times speedup does mean a lot for this community,” Wang says in explaining the interest in the paper, of which he is the chief author.
Model-based image reconstruction also allows a detailed image to be obtained with a much shorter scan, reducing radiation exposure by about three quarters of the original amount.
The Purdue team’s patented technology is in the processing of being commercialized by High Performance Imaging, a company co-founded by Bouman, a professor in Purdue's School of Electrical and Computer Engineering and Weldon School of Biomedical Engineering, and Midkiff, who likewise is an electrical and computer engineering professor, along with colleagues Anand Raghunathan, professor of electrical and computer engineering, and Sherman Kisner, a research scientist in Purdue's College of Engineering.
The company is aimed not only at CT scanning but a variety of other industrial and scientific imaging methods, from electron microscopes to synchrotrons, that collect vast amounts of data in probing biological and non-biological materials and require heavy computation to turn it into an image.
CT scans work by capturing layers, or “slices,” of a sample and reconstructing them into a 3-D image in a well-developed computational process. What isn’t as well developed is the process of imaging the slices in the first place. That’s what the technology outlined in the SC15 paper by Wang addresses.
In 2-D, digital images are made up of tiny dots called pixels. In 3-D imaging, the analogous unit is the voxel. Part of the acceleration in the Purdue technology is accomplished by grouping voxels near each other into “super voxels” and parceling those groups out to multi-core processors for processing simultaneously. The technology also anticipates what data the computer will need next and pre-fetches that into a fast memory cache so it can be loaded rapidly.
Just 20 processing cores yielded the 187 times speedup noted in Wang’s paper. That number is attainable even in a high-end workstation today, while high-performance computing clusters like those at Purdue can make hundreds or thousands of cores available.
Wang used Purdue’s new Rice cluster, which has 20 cores per node and more than 11,000 in all. Rice and the other Community Cluster Program systems offer ready access for Purdue researchers who partner with ITaP in the program, along with expert assistance available from ITaP Research Computing staff.
“If there's any trouble I have, I can find people to help me with it very quickly,” Wang says.