Brief Description
The Motion-compensation OSEM* List-mode Algorithm for Resolution-recovery Reconstruction (MOLAR) system is the result of an ongoing collaboration between three organizations in the NIH Intramural Research Program as well as Yale University, SUNY Buffalo, and CPS Innovations, Knoxville, TN, USA. It is a complete system for managing and performing high-resolution, iterative reconstructions of positron emission tomography (PET) data. MOLAR has been designed for use with the ECAT HRRT (High Resolution Research Tomograph, CPS Innovations) operating in list-mode. Due to the pluggable component design of the software, however, the MOLAR reconstruction engine is readily adaptable to any PET scanner, including frame-mode scanners. Reconstructions are performed on a parallel cluster of commodity computers. One of the goals of the project is to provide complete reference implementations (i.e., with physical effects incorporated into the model) of common iterative reconstruction algorithms such as OSEM*. Another goal of the project is to provide to the PET research community a general software framework for performing list-mode or frame-mode reconstructions on the HRRT or any other PET scanner. The framework has been designed to allow collaborating groups or individuals the opportunity to contribute their own components.
*OSEM: Ordered Subsets Expectation Maximization reconstruction algorithm.
Caption: Front view of the ECAT HRRT scanner by CPS Innovations, Knoxville, TN. The topology of the HRRT is an octagon of dual-layer (LSO/LYSO) detector banks, each bank being an array of 9*13 blocks, and each block containing 8*8 crystals of size 2.1*2.1*10-mm each. With 119,808 crystals, the HRRT has 4.5 billion potential lines of response. The HRRT has a 35-cm patient port, making it suitable for human brain studies as well as large animal studies. The massive amount of data generated by the HRRT, coupled with interest in high-resolution, high-sensitivity reconstructions, motivates our present work.
Caption: One slice of a sinogram, binned by MOLAR, of a uniformity cylinder phantom. The diamond patterns are due to the detector gaps.
Caption: Central transverse slice of the global sensitivity image generated by MOLAR with 50-M randomized events, without attenuation or motion correction. The spoke-and-wheel pattern is a result of gap-effect cancellation along the lines of response that connect two gap regions.
List of Collaborators:
• Robert Innis, M.D., Ph.D., Chief, Laboratory of Molecular Imaging, National Institute of Mental Health (NIMH)
• Jeih-San Liow, Ph.D., Laboratory of Molecular Imaging, NIMH,
• Peter Herscovitch, M.D., Director, Positron Emission Tomography Department, Clinical Center (CC)
• Craig Barker, Ph.D., Positron Emission Tomography Department, CC
• Shanthalaxmi Thada, Positron Emission Tomography Department, Clinical Center
•
Richard
Carson, Ph.D.,
•
Rutao
Yao, Ph.D.,
Major Accomplishments of this Activity in FY 2007
In 2007, a problem was discovered in the software module and algorithm used to find the location if an event in the large list-mode data file. The old algorithm occasionally would fail. The new algorithm checks for timing problems in the list-mode file, and if a problem is found, the algorithm applies a correction. It is more far more reliable and efficient.
Significant progress was made in re-engineering the software module that generates event “packets” to work on many processors. This module, called the “House”, has become a processing bottleneck due to the increasingly large data sets and the efficient parallelization of the main body of the reconstruction engine. Preliminary testing of the parallel House has indicated nearly linear speedup, even on large numbers of processors.
Anticipated Major Accomplishments of this Activity in FY 2008:
The work on parallelizing the House module needs to continue and should be completed in FY 2008. A number of algorithmic improvements from Dr. Carson and Dr. Yao will be incorporated into the core software. A concern has been raised about the algorithm that “masks” the global sensitivity image. Modifications to this algorithm to treat only the available events (as opposed to all events) need to be incorporated into the software.
Metrics
Number of users: 8
Number of successful reconstructions: 4720
Average number of processors per reconstruction: 16
Average wall-clock time per reconstruction: 5.1
Cluster Utilization: 85%
Publication in FY 2007:
M. Rodriguez, J.-S. Liow, S. Thada, M. Sibomana, S. Chelikani, T. Mulnix, C.A. Johnson, C. Michel, W.C. Barker, and R.E. Carson, “Count-Rate Dependent Component-Based Normalization for the HRRT”, IEEE Transactions on Nuclear Science, vol. 54, no. 3, pp. 486-495 (2007).