Process-Level Parallelization of Spatial Referencing
Justin M. Lapre
M.S. Thesis, Department of Computer Science, Rensselaer
Polytechnic Institute, May 2005.
ABSTRACT
Laser retinal surgery is often used to treat vision-threatening
diseases such as diabetic retinopathy and Age-related Macular
Degeneration (AMD). The failure rate for such surgeries is
approximately 50\% \cite{bressler, fine, krauss, monahan, monahan2,
murphy}, both for the initial surgery as well as for each follow-up
surgery. The spatial referencing algorithm, developed by Lin
\textit{et al.} \cite{lin} was developed to aid the physician
responsible for those surgeries. Spatial referencing allows the
surgeon to pinpoint the area over which the laser is aimed, thereby
minimizing any errors during the procedure and decreasing the rate of
failure.
In turn, the work of O'Neil \cite{oneil} tried to further enhance the
algorithm by parallelizing it. While his approach with two threads
had some positive speedup, his approach with four threads suffered a
drastic slowdown, ultimately performing worse than the
non-parallelized version of the same code.
The aim of this work is to improve upon Lin's algorithm, again by
attempting to parallelize it, though taking a different approach than
O'Neil. Our approach prefers \textit{process-level} parallelism as
opposed to \textit{thread-level} parallelism for reasons that will be
made clear in the coming chapters.
Additionally, some degree of a real-time deadline was imposed in this
code, while in most of the cited literature it is not taken into
consideration. Instead of a time-based deadline, however, we have
opted for a computational deadline.
This thesis is motivated by the need to build real-time tools to aid
physicians performing laser retinal surgery.