Positron Emission Tomography » Comput. Issues

Our MAP PET images are computed using a preconditioned conjugate gradient method. Computational costs are reduced using sparse matrix structures, run-length encoding, and automated indexing for forward and backprojection. Nevertheless, computing costs remain high for a single processor machine. For this reason, we have written multithreaded versions of our code that run under the Sun Solaris and Windows-NT operating systems. Multithreading of forward and backprojection gives close to N-fold speed up for a N-processor system (e.g. for our four processor systems we achieve a speed up by a factor of about 3.5). All code is written in C, can directly read standard CTI data formats and can be run through a motif-based graphical user interface.

Table: Reconstruction complexity for multithreaded MAP reconstruction: all times are for 20 iterations on a 4x400 Mhz, Xeon Pentium II server

Our benchmark studies on SunSPARC systems indicate run times of approximately 2-3 times longer on a 4 CPU 450MHz system. We plan to produce a Linux version of this code in the near future.

Figure: The motif-based Graphical User Interface for the USC MAP reconstruction code (currently the FORE+PWLS option is not implemented).

Future Performance: While Moore's law provides some hope that run times will eventually become arbitrarily short as computing power improves, the attached figure is quite sobering! While computers are getting faster, PET machines are getting bigger (or their detectors are getting smaller) at an even faster rate. This figure shows that improvements in run time alone are not sufficient to meet the computing demands for future PET scanners. Improvements in algorithmic implementations and multiprocessor systems will also be necessary.

Figure: Illustration of the approximate order of computational complexity for 2D and 3D clinical and small animal scanners shown in comparison to ``Moore's law'', the observation that single-processor computing power doubles roughly every 18 months. The lower curve for the ECAT systems represents 2D complexity, the upper curve represents 3D complexity.

Figure: Preliminary results for compression of raw sinogram files as a function of total counts in the sinogram. Shown are the compressed sinogram sizes in bits/per sinogram element as a function of the total number of counts in a 288x288 HR+ sinogram. The different methods show use combinations of Huffman and run-length encoding schemes. For 3D systems with large numbers of sinogram elements and few average counts per bin, lossless compression may be an alternative to the use of raw list-mode data.

USC Biomedical Imaging Research Lab © 2004, designed by Dimitrios Pantazis