Video gaming chip achieves instant image registration

June 1, 2007

A video gaming chip will make radiologists faster and their diagnoses more accurate, if IBM and the Mayo Clinic have their way. An algorithm optimized to run on this chip, which is the heart of Sony's PlayStation 3, can accelerate the postprocessing needed to register 3D data sets by 50-fold, making registrations a snap for data sets obtained from different imaging modalities or exams conducted at different times.

A video gaming chip will make radiologists faster and their diagnoses more accurate, if IBM and the Mayo Clinic have their way. An algorithm optimized to run on this chip, which is the heart of Sony's PlayStation 3, can accelerate the postprocessing needed to register 3D data sets by 50-fold, making registrations a snap for data sets obtained from different imaging modalities or exams conducted at different times.

"The algorithm tries different combinations or translations, sliding and rotating the patient images until the old and new studies-the PET, MR, or whatever-match up," said Dr. Bradley Erickson, a neuroradiologist and director of the radiology informatics lab at the Mayo Clinic in Rochester, MN. "The trick is to figure out the metric."

Time is critical when registering data sets, according to Erickson's IBM cohort, Shahrokh Daijava.

"If you start from the wrong point, the data sets might not converge to the right solution," said Daijava, IBM's software lead for next-generation systems. "If every (attempt to converge) takes five minutes and several attempts are necessary, the radiologist will get frustrated and give up."

The alignment of images sets the framework for the fusion of images, which simplifies the detection of tumor changes that may indicate the likelihood of malignancy or patient response to chemo- or radiotherapy. Such registrations are constrained by the time needed to do them, particularly since computers do not always succeed on the first try.

A test run of data sets acquired at the Mayo Clinic showed that registering key reference points from two data sets can be accomplished in less than a second, while it takes more than a minute on a conventional workstation, according to Erickson.

Two volumetric data sets can be entirely registered in less than 20 seconds compared with several minutes, he said. Registrations performed on 98 data sets using the souped-up processor and algorithm took just eight minutes and 36 seconds, compared with seven hours on a conventional processor.

The algorithm runs on an IBM BladeCenter QS20 outfitted with a Cell Broadband Engine (CBE) processor developed as part of a joint venture by IBM, Toshiba, and Sony Group. This chip was designed to handle the kind of rapid image rendering demanded by high-performance video games, but it can be used for other kinds of advanced imaging. Together, the chip and algorithm find the best spatial positioning for registering specific 2D images, translate this to all the slices in the data sets, and then very rapidly register the entire data sets.

Erickson uses the combination in a dedicated workstation to perform research at Mayo on multiple sclerosis and brain tumors, registering data sets so images can be compared exactly over time and among modalities. Future applications of the technology could be distributed across a network, using cell processors integrated into a thin-client server or even a PACS connecting desktop PCs.

"If this is properly developed as a product, you could be looking at an old and new study, click a button and-in a subsecond-that old study would be aligned to match the new one," he said.

IBM is now negotiating with several imaging companies that might bring the algorithm and Cell-powered computing hardware to the marketplace.

"IBM needs to partner with a big vendor like GE or Hitachi or Siemens that has been in this field for a long time," Daijava said. "We are negotiating with more than one. But in the end, it will be just one partner."