M.Sc. thesis defense: Graphics Hardware Acceleration Iterative Reconstruction for SPECT Imaging
Presenter: Ken Chidlow
Making medical images of the inside of the human body from a set of images from the outside of the body requires a process called tomographic reconstruction. There is a common algorithm for reconstruction, Filtered Back Projection (FBP), that works very well for X-ray Computed Tomography, but it doesn't work so well for emission tomography. In our work, we implement an iterative algorithm, Maximum Likelihood Expectation Maximization (EM), that works well for emission tomography using graphics hardware to reduce the reconstruction time. The EM algorithm is usually very computationally expensive, to the point that it is not commonly used clinically. Our implementation of the EM algorithm as well as the related Ordered Subset EM (OSEM) algorithm uses the texture mapping capabilities of the graphics hardware to handle the data. We achieved a nine fold speed up over an optimized software implementation. This result is accompanied by a large reduction in RAM required, making the hardware version more scalable. This improvement is significant as it will make this algorithm practical for clinical use.
In order to achieve a large speed up, we present a novel bit splitting method that divides the data over different color channels as an accumulation strategy. An error analysis demonstrates that our method is superior over previous bit splitting methods in terms of accuracy. The hardware based implementation also incorporates attenuation correction for improved accuracy with only a small decrease in speed.