Super-resolution PET image reconstruction with sparse

representation
Naggov Naggayev 1

1
Master student at Azerbaijan State Oil and Industrial University, 16/21 Azadliq pr, Baku, Azerbaijan.

Abstract
Reconstructed pictures in positron emission tomography (PET) research are frequently noisy
and low quality. The low-count issue is the main source of these issues. Sparse prediction is
more likely to be chosen as the solution as sparse technology is employed more frequently. I
suggest a brand-new sparse prior technique in this research to process low quality PET
reconstructed pictures. Two dictionaries (D1 for low-resolution PET images and D2 for high-
resolution PET images) are trained from plenty of actual PET image patches in the proposed
approach. The sparse representation for each patch of the input PET picture is then obtained
using D1. Finally, D2 is used to create a high-resolution PET image from this sparse
representation. The results of the studies show that the suggested strategy has superior effects
that improve image resolution and detail recovery that are stable. In terms of root mean square
error, this technique performs better quantitatively than older methods (RMSE). The
suggested method offers a fresh and effective way to enhance the image quality of PET
reconstructions.

Keywords 1
PET imaging, data acquisition, artifacts, resolution, information technologies.

1. Introduction equipment costs. In

this paper, a new idea
is proposed for PET
In recent years,
image reconstruction
positron emission
that incorporates prior
tomography (PET)
information. Edge-
has been widely
preserving
studied by researchers
regularization is the
for diagnosing
most common way to
diseases. However,
achieve good results.
low-count PET
By evaluating
projection data is a
different
continual challenge
regularization
that results in noisy
methods that use
images and generates
spatial prior
artifacts.
information, we
Many methods
found that Bicubic
have been proposed
interpolation [1],
to improve the quality
which uses selected
of the PET images.
adjacent pixel
The traditional
methods for each
reconstruction
pixel in the prior
algorithms typically
image, surpasses
have high
others in terms of
computational and
performance and
1
EMAIL: nnaggayev@cbmes.com;
computational the sparse
complexity. Another representation of a
class of methods used low-resolution PET
to incorporate prior image input based on
information are 𝐷1 would be used to
related to post- obtain a high-
processing the resolution PET image
reconstructed images. from the 𝐷2
The highly dictionary, based on
constrained back the linear
projection method relationships between
synthesizes images as the input and output
prior information. data.
The back projection The suggested
method was algorithm performs
developed for image adaptive selection
reconstruction of well, which
under sampled data in contributes to its
magnetic resonance improved capacity to
imaging (MRI) and enhance resolution
has been applied to and denoising when
denoise dynamic PET compared to the
images. regularization-based
Inspired by the methods previously
success of image stated. This work
super-resolution demonstrates how
through sparse low-count PET
representation [2], we reconstruction images
propose a similar can be enhanced by
method to obtain using sparse prior
high-resolution PET information. It is a
images by using cutting-edge strategy
appropriate sparse addressing the low-
representations of count problem in the
low-resolution PET field of PET.
images. Because the It is well
reconstruction recognized that the
constraint is not SR image
unique and the reconstruction
measurements cause problem is poorly
errors, this is a posed due to a lack of
severely ill-posed LR images and
problem. reconstruction
Consequently, we use limitations. The
regularization aforementioned
methods to solve the techniques necessitate
problem. Therefore, massive databases of
this method relies on HR and LR patch
two coupled pairings, which
dictionaries 𝐷1, necessitates a lot of
trained on low- calculations. In order
resolution PET image to address this issue,
data, and 𝐷2, trained regularization
on high-resolution techniques are used in
image data. Finally, this paper. 𝐷1 and 𝐷2
are two connected
dictionaries that are
where 𝑎 is the sparse
used in this strategy.
representation of
𝐷1 and 𝐷2 are trained
patch x for the input
using LR PET image
image Y. In this idea,
data and HR image
the sparse condition
data, respectively.
(3) is used to obtain
Both dictionaries
the representation of
have too many
each local patch.
entries.
Then, using constraint
(1) as the premise of
2.1. this step, the
Methodology representative results
are used as the basis
for recovering the
We model the
entire image.
problem as follows,
Therefore, there are
based on [1]. Given a
two models in the
low-resolution PET
method, a local model
image Y, generate a
and a global model,
high-resolution PET
Using the local
image X that exhibits
model, we train two
the same conditions.
We adopt two dictionaries and 𝐷1,
constraints in this which represent the
work to solve this textures of the images
problem. First, for the and cause the low-
sparse reconstruction, and high-resolution
the high-resolution PET images to have
image X should be the same sparse
consistent with the representation. Then,
low-resolution image according to 𝐷1, the
Y: process of finding the
sparsest
Y  SH X (1) representation can be
depicted as follows
In the preceding
equation, H is a min"𝑎"0 s. 𝑡.
sparse matrix, and S "𝐹𝐷1 − 𝐹𝑦"2 ≤ c
is the sampling (3)
matrix. Second, it
reflects the sparsity where F is the feature
prior. The high- extraction operator of
resolution PET image the low-resolution
X is segmented into a PET image. F is
series of small equivalent to a high-
patches, x, that can pass filter because the
be sparsely finer elements of an
represented in terms image frequently
of the 𝐷2 dictionary include its high-
as follows: frequency content.
The function of F is
x ≈ 𝐷 2𝑎 for some to ensure that the 𝑎
𝑎 ∈ 𝑅 𝐾 with "𝑎"0 « calculated in (3) is
𝐾 (2) closely related to the
image to be
recovered, thus is the most crucial
making the prediction stage (3). In this
more reasonable. The paper, we trained the
global model from dictionaries 𝐷 1 and
constraint (1) is used 𝐷 2, which are
to guarantee the connected. The
robustness of the corresponding low-
recovered image, and and high-resolution
it can also suppress images would have
possible artifacts that about the same sparse
arise from the local representation thanks
model. to these two
In the model, the core dictionaries.
of the problem is 2.2. Dictionary
finding the sparse
representation of the training
image. The most Three phases
significant stage is to make up the
obtain an appropriate, suggested algorithm.
overcomplete A collection of
dictionary that dictionaries LR and
matches your needs HR picture pairs are
(3). In this paper, we trained during the
trained two related first stage, which is
dictionaries 𝐷1, and called the training
𝐷2, with given PET stage. The second
image patch pairs ({𝑇 stage involves
|𝑇 = 𝑦1, 𝑦 2 , ⋯, 𝑦𝑛} sparsely
and {X |X = 𝑥1, 𝑥 2 , reconstructing the HR
⋯, 𝑥}), which can be patches from their
formulated as associated LR patches
follows: by choosing the
optimal dictionary
𝐷c = arg min "Xc − combination. The
𝐷c𝑍"2 + 𝜆"𝑍"1, third stage is the
(4) iterative SR
where 𝐷c = {𝐷1, 𝐷2}, reconstruction stage,
Xc = {𝑌, X}, and the where each HR image
that has been rebuilt
𝘗1 norm "𝑍"1 is
is sent into the stage
present to enforce
after it to improve the
sparsity. The
image quality. Fig. 1
parameter 𝜆 is
depicts the suggested
intended to balance
SR algorithm's
the sparsity of the
flowchart:
solution.
Identifying the
image's sparse
representation is the
main challenge with
these models. Getting
a proper
overcomplete
dictionary that fits Eq
 sparse coding of the
LR patches both use
these obtained
features. Finally, pairs
of LR and HR patches
are handled once they
correspond to the
same spatial
locations. The K-
singular value
decomposition (K-
SVD) approach is
used to train the
dictionaries to
produce sparse signal
representations that
better fit the data.
An LR image is
initially magnified
using bicubic
interpolation to have
the same dimensions
as the equivalent HR
image at the SR
Figure 1
reconstruction step,
A set of HR images which is identical to
and a set of LR the training stage.
images are needed for Applying feature
the training stage. The extraction filters
related HR images are allows for the
down - sampled and extraction of features,
blurred using an which are
average filter to subsequently molded
create the LR images. into vector forms. We
Additionally, using an apply a certain patch
interpolation process, overlap with specified
the LR images are pixels in order to
enlarged to have the guarantee local
same dimensions as consistency between
the corresponding HR the reconstructed
images. A feature patches.
extraction filter is
We initially
applied to the
determine the sparse
interpolated LR
representation
pictures to extract
coefficient vector of
features. Due to their
the appropriate HR
efficiency and
feature vector across
simplicity, we adopt
the cluster LR
the first- and second-
dictionary using the
order derivatives in
dictionary pair of the
this article as the
discovered cluster.
feature for the LR
Following that, the
patch. The dictionary
HR patch is rebuilt by
learning process and
right multiplying the
cluster HR dictionary PET images, we
by the derived sparse chose 3 × 3-pixel
representation patches, while for the
coefficient vector. By high-resolution
integrating the two- dictionary, we chose
dimensionally 9 × 9-pixel patches
reshaped and from high-resolution
reconstructed HR PET images.
patches, an HR image We demonstrate
is finally produced. the effectiveness of
the suggested SR
Experimental approach in this
section. For each
Results dictionary, we choose
1000 photos of the
To observe the highest calibre. All of
performance of the the LR images used in
proposed method, we the training or testing
use an emission phase were created by
activity phantom a simply down
simulated data sampling and blurring
experiment as shown the original high-
in Fig. 2 (a). The resolution HR
resolution of each photographs. The
image is 256×256. down sampled and
The low-resolution blurred test images
image is the result of have noise applied to
the PET iterative them. Prior to
reconstruction that processing the LR
used the maximum image, 100,000 patch
likely hood pairings randomly
expectation picked from CT
maximization images gathered from
(MLEM) algorithm. clinical hospitals and
The bicubic phantom simulations
interpolation were used to train two
algorithm was been dictionaries (Dl and
utilized and compared Dh). We chose a
with our method. dictionary size of
Before processing the 2048 atoms to
low-resolution image, achieve a compromise
we trained two between computation
dictionaries 𝐷1 and 𝐷2 and image quality;
from 100,000 patch this number of atoms
pairs randomly can fully describe the
sampled from PET aspects of the photos
images collected from without being too
medical image computationally
libraries on the demanding for our
internet. To limit the computer setup.
time spent for training Empirically, we
the dictionary, we chose 5 5-pixel
fixed its size to 1024 patches with an
in our experiment. overlap of 4 pixels
For the low-resolution between them and a
sparsity regularization zoomed area of the
of = 0.4. In this study, image. Compared
the proposed to (b), image (c),
approach was which was
assessed using head recovered by
data from clinical bicubic
CBCT and XCAT
interpolation, has
phantom simulation.
an obvious
To assess the output
SR image quality, we improvement. In
employed the metrics (c), the edges have
RMSE, PSNR, and become clear and
SSIM. the detail of the
zoomed area has
been recovered.
Simultaneously,
the noise is
reduced, but the
bicubic algorithm's
performance in
suppressing
artifacts is
unsatisfactory.
Although image (c)
has a better quality
than image (b),
the edges in the
zoomed area in
(g) are slightly
blurrier than the
edges in (h). We
also provide the
RMSE values for
the images
compared to the
Fig. 2. Results with a true image (a) in
brain phantom. (a) Original
normal-resolution image; Table. The
proposed method
(a) the low-resolution achieved the
image; (c) Output
processed by bicubic lowest RMSE
interpolation; score.
(d) Output processed by
our method; (e)–(h) are the
zoomed areas of (a)– (d

In Fig. 2, the
low-resolution
image (b) is 2. Conclusions
blurry and
contains artifacts, A single-CT
especially in the image SR
reconstruction hot field for
approach is researchers. The
suggested in this application of
research. This SR sparse prior
reconstruction information in
method is based on processing the PET
dictionary training images seems to be
with a collection of a potential strategy
LR and HR picture for solving the
patch pairs and problems caused
sparse by low-count PET
representation image
theory. The applied reconstruction;
sparse domain's specifically, better
ability to image
accurately reconstruction
represent the target means that PET
image by jointly scan times and
training the LR and radiation doses can
HR dictionaries be reduced. In this
has a significant study, a super-
impact on the resolution method
quality of the SR was proposed for
reconstruction PET image
images. We create
a sparse 3. Acknowledge
representation for
each patch of LR
ments
photos and make
An
sure that the local
acknowledgment
patches in LR and
section, which goes
HR are similar. just before the
The HR reference section in
counterpart is your document,
rebuilt using the should list all of the
coefficients from funding sources and
the LR domain. other sources of
We incorporate a assistance that were
number of iteration used as well as
methods express gratitude to
throughout the HR the people and
image organizations that
helped with the
reconstruction
research and
process to enhance production of the
image quality. work. Aleksandr
Due to the poor Ometov, a TAU
quality of low- employee in Finland,
quality PET developed this Word
template. The
images, recovery
Creative Commons
work has become a
Attribution-Share
Alike 4.0 [5] Trinh DH, et al.
International license Novel Example-
is used to distribute Based Method
the template (CC BY- for Super-
SA 4.0). Resolution and
Denoising of
Medical
Images. Ieee T
Image
4. References Process. 2014.
[6] Sun J, Sun J, Xu
ZB, Shum HY.
[1] Sun J, Sun J, Xu
Image super-
ZB, Shum HY.
resolution using
Image super-
gradient profile
resolution using
prior. 2008 Ieee
gradient profile
Conference on
prior. 2008 Ieee
Computer Vision
Conference on
and Pattern
Computer Vision
Recognition. 200
and Pattern
8.
Recognition. 200
[7] Dong WS,
8.
Zhang L, Shi
[2] Zhou F, Yang W,
GM, Wu XL.
Liao Q.
Image
Interpolation-
Deblurring and
Based Image
Super-Resolution
Super-Resolution
by Adaptive
Using
Sparse Domain
Multisurface
Selection and
Fitting. IEEE
Adaptive
Transactions on
Regularization. I
Image
eee T Image
Processing A
Process. 2011.
Publication of
[8] Lu Y, Zhao J,
the IEEE Signal
Wang G. Fair-
Processing
view image
Society. 2012
reconstruction
[3] Farsiu S,
with dual
Robinson MD,
dictionaries. Phy
Elad M, Milanfar
sics in Medicine
P. Fast and
& Biology. 2012.
robust
[9] Ji DJ, Qu GR,
multiframe super
Liu BD.
resolution. Ieee T
Simultaneous
Image
algebraic
Process. 2004.
reconstruction
[4] Lin, Z. & Shum,
technique based
H. Y.
on guided image
in Computer
filtering. Optics
Vision and
Express. 2016.
Pattern
[10] Freeman WT,
Recognition,
Jones TR,
2001
Pasztor EC.
Example-based
 super-resolution.  CT for the
IEEE Computer Diagnosis of
graphics and Acute
Applications. 20 Appendicitis: A
02. Meta-Analysis. 
[11] Aharon M, Elad Am J
M, Bruckstein A. Roentgenol. 201
K-SVD: An 7.
Algorithm for
Designing
Overcomplete
Dictionaries for
Sparse
Representation. I
EEE
Transactions on
Signal
Processing. 2006
.
[12] Segars WP,
Sturgeon G,
Mendonca S,
Grimes J, Tsui
BMW. 4D
XCAT phantom
for
multimodality
imaging
research. Medica
l Physics. 2010.
[13] Donoho DL. For
most large
underdetermined
systems of linear
equations, the
minimal. Comm
unications on
pure and applied
mathematics. 20
06.
[14] Tibshirani, R.
Regression
shrinkage and
selection via the
lasso. Journal of
the Royal
Statistical
Society. Series B
(Methodological
), 2006.
[15] Yun SJ, et al.
Comparison of
Low- and
Standard-Dose