% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@ARTICLE{Xu:863488,
author = {Xu, Hancong and Lenz, Mirjam and Caldeira, Liliana and Bo,
Ma and Pietrzyk, Uwe and Lerche, Christoph W and Shah, Nadim
Jon and Scheins, Juergen J},
title = {{R}esolution modeling in projection space using a
factorized multi-block detector response function for {PET}
image reconstruction},
journal = {Physics in medicine and biology},
volume = {64},
number = {14},
issn = {1361-6560},
address = {Bristol},
publisher = {IOP Publ.},
reportid = {FZJ-2019-03542},
pages = {145012},
year = {2019},
abstract = {Positron emission tomography (PET) images usually suffer
from limited resolution and statistical uncertainties.
However, a technique known as resolution modeling (RM) can
be used to improve image quality by accurately modeling the
system's detection process within the iterative
reconstruction. In this study, we present an accurate RM
method in projection space based on a simulated multi-block
detector response function (DRF) and evaluate it on the
Siemens hybrid MR-BrainPET system. The DRF is obtained using
GATE simulations that consider nearly all the possible
annihilation photons from the field-of-view (FOV).
Intrinsically, the multi-block DRF allows the block
crosstalk to be modeled. The RM blurring kernel is further
generated by factorizing the blurring matrix of one
line-of-response (LOR) into two independent detector
responses, which can then be addressed with the DRF. Such a
kernel is shift-variant in 4D projection space without any
distance or angle compression, and is integrated into the
image reconstruction for the BrainPET insert with single
instruction multiple data (SIMD) and multi-thread support.
Evaluation of simulations and measured data demonstrate that
the reconstruction with RM yields significantly improved
resolutions and reduced mean squared error (MSE) values at
different locations of the FOV, compared with reconstruction
without RM. Furthermore, the shift-variant RM kernel models
the varying blurring intensity for different LORs due to the
depth-of-interaction (DOI) dependencies, thus avoiding
severe edge artifacts in the images. Additionally, compared
to RM in single-block mode, the multi-block mode shows
significantly improved resolution and edge recovery at
locations beyond 10 cm from the center of BrainPET insert in
the transverse plane. However, the differences have been
observed to be low for patient data between single-block and
multi-block mode RM, due to the brain size and location as
well as the geometry of the BrainPET insert. In conclusion,
the RM method proposed in this study can yield better
reconstructed images in terms of resolution and MSE value,
compared to conventional reconstruction without RM},
cin = {INM-4 / INM-11 / JARA-BRAIN / JARA-HPC},
ddc = {530},
cid = {I:(DE-Juel1)INM-4-20090406 / I:(DE-Juel1)INM-11-20170113 /
$I:(DE-82)080010_20140620$ / $I:(DE-82)080012_20140620$},
pnm = {573 - Neuroimaging (POF3-573) / Modelling Realistic
Distributions for High-Resolution Positron Emission
Tomography (PET) $(jinm40_20160501)$},
pid = {G:(DE-HGF)POF3-573 / $G:(DE-Juel1)jinm40_20160501$},
typ = {PUB:(DE-HGF)16},
pubmed = {pmid:31158824},
UT = {WOS:000475775200005},
doi = {10.1088/1361-6560/ab266b},
url = {https://juser.fz-juelich.de/record/863488},
}