A deeper view of image reconstruction

This blog has mentioned Impulse Accelerated Technologies Inc. in the past for its software work in optimizing image processing for FPGAs. Thanks to a new online article from Vision Systems Design, however, we now can get a better look at what tomographic image reconstruction means in terms of fast imaging in medical and radar fields. The article is based on a Masters’ thesis from University of Washington student Nikhil Subramanian.

I was immediately fascinated in examining the image “slices” shown in the article and thesis, since I visited the eye doctor June 22 and viewed comparative tomographics slices of images of my retina over a four-year time span. When comparative images are filtered over time to yield a composite image of what has changed in blood vessels, nerves, etc., the power of diagnosis in near-real-time is beyond breathtaking.

The tough problem the researchers at UW were trying to solve was the back-projection algorithm, used to map the data of the image into a 3D image space to reconstruct the body part (or radar image, etc.) from the acquired data. Since this is a task crying out for parallelism, running the software on an integer processor can be wasteful and time-consuming. The author studied some existing parallel graphic subsystems, such as the Annapolis Micro Systems board using a Virtex 2000E, but elected to use the XtremeData accelerator, marrying an Opteron processor with an Altera Stratix-II.

You can read the article and thesis for details, but the author discovered the key to fast image reconstruction was not so much the hardware itself, as the tools for rapid parallelization, represented in this case by the Impulse C C-to-FPGA tool. While using FPGAs with standard software could provide some advantages over traditional DSPs or integer processors, an FPGA subsystem using the pipelining of Impulse tools offers significant advantages in memory utilization for parallel algorithms like back-projection. If these results are duplicated in fields outside medical imaging, FPGAs soon could dominate 3D image reconstruction applications.