Researchers are applying supercomputing technology to improve breast cancer screening accuracy.
A team of researchers will use supercomputing with a 30-year-old imaging modality called tomosynthesis, which until now has been relegated to research labs due to its massive and expensive computational requirements.
Researchers at the Northeastern University Computer Architecture Research Lab (NUCAR) and the National Science Foundation’s (NSF) Centre for Subsurface Sensing and Imaging Systems (CenSSIS) are teaming with Massachusetts General Hospital (MGH) for the supercomputing technology.
Called Digital Breast Tomosynthesis (DBT), the system creates a 3D image of the breast using a series of X-ray projections collected during a 20-second, 40-degree sweep. It makes cancer lesions easier to detect among dense breast tissue by creating a stack of 1mm spaced high-resolution slices that can be displayed individually, or assembled into a 3D view that can be rendered for more careful examination. DBT also reduces the amount of breast compression required by traditional mammography, which can deter women from getting an annual screening.
NUCAR scientists addressed this computational hurdle by creating a DBT reconstruction application on their desktop PCs using Matlab, and then running the code on an affordable Linux parallel cluster using Star-P software from Interactive Supercomputing (ISC). Reconstructing DBT used to take many hours. With this new Star-P approach, imaging reconstruction can be completed in a few minutes. The complicated parallel programming task has been dramatically simplified using Star-P, slashing development time from many months to days.
General Electric is developing a commercial DBT device that should be generally available in 2009. The Northeastern team has been supported by a NSF Small Business Technology Transfer (STTR) project with ISC, the NSF’s CenSSIS, and by Mass General Hospital, which is known worldwide for offering the most advanced breast screening and treatment services available.
‘The support from Northeastern has radically shifted this paradigm,’ said Richard Moore, program director for Breast Imaging Research and CenSSIS at MGH. ‘With this kind of performance, we can realistically rely on 3D methods that were out of the question previously. It’s not just the speed, it’s the exploratory freedom.’
Professor David Kaeli, director of NUCAR and thrust leader in CenSSIS, said: ‘The multi-dimensional imaging technique involves the processing of up to 15 high resolution x-ray images. This kind of application typically requires a very long time to carry out repetitive operations on large image matrices. Parallelising these large datasets on affordable hardware can now achieve the performance required for real-world implementations.' The parallelisation effort was performed by Dana Schaa, a graduate student working with Professor Kaeli.
Kaeli said Star-P was a good solution for this challenge because it enabled his team of researchers to easily code algorithms using their familiar desktop environment, automatically transforming the application to run on parallel clusters. Star-P eliminates the need to re-program the applications in complex languages such as C, FORTRAN or MPI (message passing interface) to run on the cluster - which otherwise requires arcane programming knowledge and months to complete.