Speaker
Mr
Niklas Rehfeld
(Section for Biomedical Physics, Radiooncology, University of Tuebingen)
Description
Monte Carlo (MC) simulations in positron emission tomography (PET) play an important
role in detector modeling and algorithm testing. Whereas the simulations are widely
used in a forward projection manner to accomplish this task, ideally they should be
included into the reconstruction process itself. It is therefore desirable to
investigate the convergence properties and the propagation of MC noise of these kind
of reconstruction algorithms. For human scanners the correct treatment of patient
scatter plays a dominant role. The incorporation of this kind of scatter into the
matrix is therefore important.
MC simulations were integrated into the maximum likelihood expectation
maximization(ML-EM) algorithm in two different ways. In the full matrix approach the
system matrix was calculated by running MC simulations including scatter. This matrix
was used in both the projector and the back-projector. In the dual matrix (DM)
approach, MC simulations were used to incorporate scatter in the projector, whereas
the back-projector only comprised attenuation. Repeated reconstructions with
different MC seeds allowed a statistical analysis of the error at each iteration step
and made it possible to investigate separately the propagation of the MC noise that
was introduced by the sinogram, by the projector, and by the matrix.
Both approaches resulted in similar images, but the DM approach with unmatched
projector and back-projector yielded a faster initial convergence and a faster
divergence at higher iteration numbers when compared to the ideal full matrix
approach. The analysis of the noise sources for the modeled 2D-scanner in full matrix
reconstruction showed that the noise introduced by the matrix became comparable to
the noise introduced by the sinogram when using a matrix that was simulated with
10000 emissions/voxel or less using variance reduction techniques. The time needed to
simulate this matrix with 6400 voxels and 234700800 elements was less than four
minutes on a small computer cluster with eight two-processor computers.
Author
Mr
Niklas Rehfeld
(Section for Biomedical Physics, Radiooncology, University of Tuebingen)
Co-author
Dr
Markus Alber
(Section for Biomedical Physics, Radiooncology, University of Tuebingen)
