Research Article ISSN 2771-9014

Research Article Journal of Biotechnology and its Applications

Sinogram Interpolation Inspired by Single-Image Super Resolution

Carolyn Christiansen1 and Gengsheng L. Zeng1,2*

*
Correspondence:
Department of Computer Science, Utah Valley University,
1

Orem, Utah, USA. Gengsheng L. Zeng, Department of Computer Science, Utah

Valley University, Department of Radiology and Imaging
Department of Radiology and Imaging Sciences, University of
2 Sciences, University of Utah, Salt Lake City, Utah, USA.
Utah, Salt Lake City, Utah, USA.
Received: 04 Apr 2023; Accepted: 09 May 2023; Published: 15 May 2023

Citation: Christiansen C, Zeng GL. Sinogram Interpolation Inspired by Single-Image Super Resolution. J Biotechnology App. 2023;
2(1); 1-7.

ABSTRACT
Computed tomography is a medical imaging procedure used to estimate the interior of a patient or an object.
Radiation scans are taken at regularly spaced angles around the object, forming a sinogram. This sinogram is
then reconstructed into an image representing the contents of the object. This results in a fair amount of radiation
exposure for the patient, which increases the risk of cancer. Less radiation and fewer views, however, leads to
inferior image reconstruction. To solve this sparse-view problem, a deep-learning model is created that takes as
input a sparse sinogram and outputs a sinogram with interpolated data for additional views. The architecture
of this model is based on the super-resolution convolutional neural network. The reconstruction of model-
interpolated sinograms has less mean-squared error than the reconstruction of the sparse sinogram. It also has less
mean-squared error than a reconstruction of a sinogram interpolated using the popular bilinear image-resizing
algorithm. This model can be easily adapted to different image sizes, and its simplicity translates into efficiency in
both time and memory requirements.

Keywords a sinogram. This sinogram is then reconstructed into an image

Deep learning, Limited data imaging, Machine learning, Medical representing the contents of the object [1]. This results in a fair
imaging, Tomography. amount of radiation exposure for the patient, which increases
the risk of cancer. A successful model means less radiation can
Abbreviations be used in a scan. Better images provide better information for
CT: Computed tomography, FBP: Filtered backprojection, SISR: doctors, which can produce better decisions. This will hopefully
Single image super-resolution. translate into better patient outcomes.

Introduction More views mean more radiation exposure for the patient.
However, using current methods, the under-sampling leads to
In medical imaging, any patient exposure to radiation means
artifacts and inferior image reconstruction. Several methods of
risk. Hence, it is ideal and morally sound to limit this exposure. solving this sparse-view problem have been proposed. Algorithms
However, this limitation in exposure translates to a sparsity of operate in various domains in this endeavor (reconstructed image
data for medical scans, and potential loss in imaging information. domain or sinogram domain). Many, but not all, of the solutions
The goal of this project is to construct a reasonably detailed involve deep learning.
representation of imaged objects using sparse data.
Some methods remove artifacts in the reconstructed image.
Computed tomography (CT) is a medical imaging procedure used Outside of machine learning, an iterative algorithm [2] has been
to estimate the interior of a patient or an object. Radiation scans used. Various deep learning models applied to this problem
are taken at regularly spaced angles around the object, forming include a general adversarial network (GAN) [3], U-net model

J Biotechnology App, 2023 Volume 2 | Issue 1 | 1 of 7

[4], and a model employing wavelet transform [5]. Other methods
attempt to interpolate the sinogram. Outside of machine learning,
a sine wave approximation technique [6] has been proposed. Deep
learning models applied to this problem include a U-net model
[7], and a U-net and residual model [8]. Yet other deep learning
models work in both domains [5,9,10]. These complex methods
attempt to combine information from the sparse sinogram and the
reconstructed image.

The model presented here falls into the second category,

interpolating the sparse sinogram and reconstructing the
interpolation. As the model will be based on SISR, several
deep learning models that have been applied to generic image
Figure 1: A simple elliptical phantom.
reconstruction should be noted. A few of the current models are
the Super-Resolution Convolutional Neural Network (SRCNN)
[11], residual models including MSRN [12,13] and EDSR [14], Sinograms and Ct Image Reconstruction
and the inception model [15]. There are other ways to extend the Computed tomography (CT) is a medical imaging procedure used
to estimate the interior of a patient or an object. For parallel-beam
sparse data such analytic extension [22,23], deformation [24],
CT, one-pixel-tall radiation scans are taken at regularly spaced
nonlinear filtering [25], and so on.
angles aroundthe object (Figure 2) and then stacked on top of each
other, forming a sinogram (Figures 13-16). Applying the Central
The model attempts to solve this sparse-view problem by
Slice Theorem and a process called filtered backprojection (FBP),
interpolating the sinogram using techniques from single image
this sinogram is then reconstructed into a two-dimensional image
super-resolution (SISR). In the literature review, no papers were
representing a slice of the contents of the object, as shown in
found that use this exact method to estimate a full sinogram
Figures 9-12 and 17-20 [1]. Several slices are usually taken in the
from a sparse sinogram. Specifically for parallel-beam CT, the
same patient.
deep learning model will accept the input of a sparsely sampled
sinogram and output the prediction for a corresponding full
sinogram. The goal of the model is that predicted full sinogram
should be a reasonable reconstruction of the original image, with
the error less than that of a reconstruction of thesparse sinogram.

Methods
Phantoms
In developing and testing medical technology, simulated images
called phantoms are used instead of real patients (or patient data).
This practice is common in the field, as it provides simulated data
that can be used to test a model.

In the first stage of the project, phantoms of one or two ellipses

of varying intensities are generated on a black background [16]. Figure 2: A CT image from TCIA dataset.
Anything outside the reconstruction circle is blacked out to allow
direct comparison of the phantom with the model predicted It should be noted that the number of views in a true, “gold-
reconstruction. These represent ground truth, or the ideal expected standard” sinogram will be much greater in clinical practice.
result. There are 1000 images of size 128 × 128 area generated. These numbers are chosen for ease in viewing artifacts. According
These are used to train the model. There are 250 additional images to Zeng [20], for a scan using 896 detectors, measurements from
that are generated to test the model. 1200 views over 360° are considered as a full sinogram, and
measurements from 400 views over 360° are considered as an
In the second stage of the project, clinical reconstructed CT images under-sampled sinogram.
are downsized and used as more complex, realistic phantoms.
Clinical images are obtained from The Cancer Imaging Archive In sparse-view tomography, measurements are under-sampled,
[17-19] and are all of size 512 × 512 pixels. The 1192 images are which usually result in severe aliasing artifacts as streaking lines
randomly chosen from 298 patients and resized to 128 × 128 pixels in the reconstructed images. Hence, a sparse sinogram will contain
to use as phantoms to train the model. Another 298 are randomly 1/3 theviews of a full sinogram, or fewer. For the image in Figure
chosen to test the model. The sets of test images and training images 2, a sparse sinogram with 16 views is represented by Figure 13,
are disjoint. Sample phantoms are shown in Figures 1 and 2. and artifacts are obvious in the reconstruction (Figure 17). Success

J Biotechnology App, 2023 Volume 2 | Issue 1 | 2 of 7

of the project is measured by how closely the model-predicted case, it must grow four times in the angular (vertical) direction
reconstruction matches the original phantom, which is ground (Figure 6).
truth. In a clinical setting, however, this ground truth is impossible
to access, since an accurate picture of the patient’s interior is
unavailable. Hence, any reconstructed image will be compared
to the full-sinogram reconstruction using the mean-squared error
metric. The model cannot be reasonablyexpected to perform better
than the reconstruction of the full sinogram, which is the model’s
target. However, the goal is to produce results superior to the
image reconstructed from the sparse sinogram.

Image Resizing Algorithms Figure 5: Classic SISR: Each pixel of the input image is extended in both
Several algorithms exist for resizing an image. The bilinear dimensions.
algorithm (called linear throughout the paper) is one of these. It
uses linear approximation in two dimensions to approximate pixel
values when resizing an image. The model does not use this method,
but it provides a useful comparison to show the deficiencies of a
different algorithm used to solve the same problem to interpolate a
full sinogram from a sparse one.

Single Image Super-Resolution

Single image super-resolution is a process that approximates high-
definition images from low-definition images. Most models for
SISR have two main parts (shown in Figure 3): feature extraction
of the low- definition image, and reconstruction of a high-definition
object using a transposition layer (sometimes called a pixel shuffle
layer). This layer converts several feature layers into one larger
layer by rearranging the tensor (pixel) values. To illustrate, Figure
4 represents pixel shuffle from four layers to one layer. Figure 6: Sinogram interpolation: The sinogram is only extended in the
angular (vertical) dimension.

TensorFlow is a software library that focuses on training deep

Figure 3: Common architecture for SISR.
neural networks. Using the Python application-programming
interface Keras to access TensorFlow functionality, a deep-learning
model is constructed. As inputs to the model, “sparse” sinograms
with 16 views are created by applying Python’s radon() function
to the training images (Figure 13). As targets of the model, “full”
sinograms with 64 views are created by applying Python’s radon()
function to the training phantoms (Figure 16).

After trying many model architectures for infilling sparse

sinograms, the best results have come from a few simple dense
2D Convolutional layers (a super-resolution convolutional neural
Figure 4: Visual representation of pixelshuffle. network [11]) followed by a modified pixelshuffle tail, as shown in
Figure 7. All kernel sizes are 3 × 3. This modified pixelshuffle tail is a
Deep-Learning Model Architecture custom transposition layer. At the end of feature extraction, each pixel
The sparse sinograms contain 1/4 of the views of the full sinograms. of the input image corresponds to four pixels, one on each feature
The project goal is to approximate the full sinograms from the layer. These four pixels must be transposed onto one layer, stacked
sparse sinograms. Sinogram approximation closely resembles the vertically, to form the output-interpolated sinogram (Figure 8).
goal of single-image super resolution (SISR). SISR aims to grow
the image in both dimensions, maintaining the aspect ratio of the The random generation of initial model weights introduces
low-definition image (Figure 5). This sinogram approximation uncertainties into this process. Another experimenter using the
aims only to expand theimage in one dimension. In this particular same code and data may produce a slightly different model.

J Biotechnology App, 2023 Volume 2 | Issue 1 | 3 of 7

 Figure 10 shows the sparse view reconstruction using the liner
interpolation method to extend the sparse view into full view.
Figure 11 shows the sparse view reconstruction using the proposed
method to extend the sparse view into full view. Figure 12 shows
the full view reconstruction with all measured data. The full view
data consists of 128 views and the sparse view data consists of 16
views.

On test data, the average mean-squared error (MSE) differences

between reconstructions and fullreconstruction for simple ellipses
phantoms are:
• 3.67 for model-predicted sinogram reconstruction.
• 17.70 for sparse sinogram reconstruction.
• 13.92 for linear image-resize sinogram reconstructions.

The model-reconstruction error is twenty-one percent of the sparse-

reconstruction error. By comparison, the linear-reconstruction
error is seventy-nine percent of the sparse-reconstruction error. The
model results are superior to those from using linear interpolation.

Results from the clinically measured data are shown in Figures

13-20. When the model is trained on more complex clinical
phantoms, Figure 17 is the reconstruction of a sparse sinogram.
Figure 7: Model architecture.
Artifacts are visible as accordion-like gray lines in the image.
The reconstruction of a linear-interpolated sonogram (Figure 18)
demonstrates rotational tendencies. The reconstruction of the
model-interpolated sinogram (Figure 19) shows a decrease in
artifacts. The most ideal, full sinogram reconstruction is shown in
Figure 20. The associated sinogram-domain data is displayed in
Figures 13-16.

On clinical test data, the average mean-squared error (MSE)

differences between reconstructions andfull reconstruction are:
• 60.80 for model-predicted sinogram reconstruction.
Figure 8: Model architecture demonstrating custom transposition layer.
• 152.56 for sparse sinogram reconstruction.
• 126.67 for linear image-resize sinogram reconstructions.
Model Advantages
This model is simple, as compared to many SISR models. This
The model-reconstruction error is forty percent of the sparse-
translates into efficiency in time and memory space. From the
reconstruction error. By comparison, the linear-reconstruction
author’s experience, this model is generally faster than a residual
error is eighty-three percent of the sparse reconstruction error. The
model. It also may occupy less space in memory than a model
that upsizes the sparse sinogram before applying filters and other model results aresuperior to those from using linear interpolation.
feature extraction strategies.
The error for the model-predicted reconstructions is less than error
A strong advantage of this model is the flexibility of input and for the sparse sinogram reconstructions. The error for the model-
output sizes. The input sparse sinogram may be any size, and the predicated reconstructions is also less than error for reconstructions
interpolated output full sinogram may be any size that is a multiple from linear-resized sinograms. The model can be declared a
ofthe input size. Dimensions are not constrained to powers of 2, as relative success. Other sparsities and views with 128 collectors.
in some u-net models, and the input sinogram does not need to be Using the same downsized clinical images as phantoms, similar
resized before the model can accept it. models are trained and tested. Only the number of views and the
sparsity (ratio between sparse views and full views) is changed.
Results MSE values are recorded for reconstructions of test phantoms,
The computer simulated elliptical phantom studies are shown using the full reconstruction as truth. Results are shown in Table 1.
in Figures 9-12. Figure 9 shows the sparse view reconstruction.

J Biotechnology App, 2023 Volume 2 | Issue 1 | 4 of 7

Table 1: MSE for models of varying views and sparsities.
Model MSE as fraction Linear MSE as fraction
Sparse views Full Views
of sparse MSE of sparse MSE
128 256 0.70 1.76
85 256 0.61 2.62
64 256 0.51 1.40
64 128 0.56 1.20
43 129 0.47 1.13
32 128 0.41 1.01
32 64 0.49 0.98
21 63 0.44 0.93
16 64 0.40 0.83

Figure 12: Reconstruction from full sinogram of elliptical phantom.

Figure 13: Sparse sinogram of downsized clinical image: 16 views, 128

collectors.

Figure 9: Reconstruction from sparse sinogram of elliptical phantom.

Figure 14: Linear-resized sinogram of downsized clinical image: 64

views, 128 collectors.

Figure 10: Reconstruction from linear-interpolated sinogram of elliptical

phantom.

Figure 15: Model-interpolated sinogram of downsized clinical image: 64

views, 128 collectors.

Figure 16: Full (ideal) sinogram of downsized clinical image: 64 views,

Figure 11: Reconstruction from model-interpolated sinogram of elliptical 128 collectors.
phantom.

J Biotechnology App, 2023 Volume 2 | Issue 1 | 5 of 7

 Conclusions
A deep-learning model based on SISR can be trained to interpolate
a full sinogram from a sparse sinogram, reducing the error of the
reconstructed image. A successful implementation of this model
in a clinical setting will decrease patient radiation exposure and
reduce associated health consequences, whilestill providing a high
quality reconstructed image.

This model has room for improvement. Reconstructions from

outputs of this model appear a little blurred when compared with
reconstructions from a full sinogram. Adding a second part of the
model that works in the reconstruction domain, as in “High quality
Figure 17: Reconstruction from sparse sonogram. imaging from sparsely sampled computed tomography data with
deep learning and wavelet transform in various domains” [5], may
improve the reconstruction of this image. This model should also
be tested on noisy sinograms. In the presence of noise, the first
layer of the model may benefit from a larger kernel size.

Availability of Data and Materials

The data that support the findings of this study are available
from The Cancer Imaging Archive but restrictions apply to the
availability of these data, which were used under license for the
current study, and so are not publicly available. Data are however
available from the authors upon reasonable request and with
permission of The Cancer Imaging Archive.

Figure 18: Reconstruction from linear-interpolated sonogram. Funding

This research work was supported in part by grant R15EB024283
from the U.S. National Institute of Health.

The LDCT dataset in TCIA was funded by grants EB017095 and

EB017185 (Cynthia McCollough, PI) from the National Institute
of Biomedical Imaging and Bioengineering.

References
1. Zeng GL. Medical Image Reconstruction: A Conceptual
Tutorial. Springer, Manhattan, New York City. 2010.
2. Huang J, Zhang Y, Ma J, et al. Iterative image reconstruction
for sparse-view CT using normal-dose image induced total
variation prior. PloS One. 2013; 8: 79709.
Figure 19: Reconstruction from model-interpolated sonogram.
3. Wang H, Wang N, Xie H, et al. Two-stage deep learning
network-based few-view image reconstruction for parallel-
beam projection tomography. Quantitative Imaging in
Medicine and Surgery. 2022; 12: 2535.
4. Han Y, Ye JC. Framing U-net via deep convolutional
framelets: Application to sparseview CT. IEEE Transactions
on Medical Imaging. 2018; 37: 1418-1429.
5. Lee D, Choi S, Kim HJ. High quality imaging from sparsely
sampled computed tomography data with deep learning and
wavelet transform in various domains. Medical physics. 2019;
46: 104-115.
6. Kalke M, Siltanen S. Sinogram interpolation method for
sparse-angle tomography. Applied Mathematics. 2014; 5:
Figure 20: Reconstruction from full sonogram. 423-441.

J Biotechnology App, 2023 Volume 2 | Issue 1 | 6 of 7

7. Lee H, Lee J, Kim H, et al. Deep-neural-network-based 15. Muhammad W, Aramvith S. Multi-scale inception based
sinogram synthesis for sparse-view CT image reconstruction. super-resolution using deep learning approach. Electronics.
IEEE Transactions on Radiation and Plasma Medical Sciences. 2019; 8: 892.
2018; 3: 109-119. 16. Paleo P. Localtomo. 2016.
8. Dong X, Vekhande S, Cao G. Sinogram interpolation for 17. McCollough C, Chen B, Holmes III DR, et al. Low Dose
sparse-view micro-CT with deep learning neural network. In: CT Image and Projection Data (LDCT-and-Projection-data)
Medical Imaging 2019: Physics of Medical Imaging. 2019; (Version 4) [Data set]. The Cancer Imaging Archive. 2020.
692-698.
18. Moen TR, Chen B, Holmes IDR, et al. Low dose CT image
9. Zhang Q, Hu Z, Jiang C, et al. Artifact removal using a and projection dataset. Medical Physics. 2021; 48: 902-911.
hybriddomain convolutional neural network for limited-
19. Clark K, Vendt B, Smith K, et al. The cancer imaging archive
angle computed tomography imaging. Physics in Medicine &
Biology. 2020; 65: 155010. (TCIA): Maintaining and operating a public information
repository. Journal of Digital Imaging. 2013; 26: 1045-1057.
10. Deng K, Sun C, Gong W, et al. A limited-view CT
reconstruction framework based on hybrid domains and 20. Zeng GL. Sparse-view tomography via displacement function
spatial correlation. Sensors. 2022; 22: 1446. interpolation. Visual Computing for Industry Biomedicine
and Art. 2019; 2: 1-10.
11. Dong C, Loy CC, He K, et al. Image super-resolution using
deep convolutional networks. IEEE transactions on pattern 21. Pixel Shuffle Super Resolution with tensorflow, Keras, and
analysis and machine intelligence. 2015; 38: 295-307. Deep Learning, Online 2021.
12. Liu Y, Zhang X, Wang S, et al. Progressive multi-scale 22. Zeng GL. Iterative analytic extension in tomographic imaging.
residual network for single image super-resolution. 2020. Visual Computing for Industry, Biomedicine and Art. 2022; 5:
4.
13. Li J, Fang F, Mei K, et al. Multi-scale residual network for
image super-resolution. In: Proceedings of the European 23. Zeng GL, Li Y. Analytic continuation and incomplete data
Conference on Computer Vision (ECCV), Munich, Germany. tomography, Journal of Radiology and Imaging. 2021; 5: 5-11.
2018; 517-532. 24. Zeng GL. Sparse-view tomography via displacement function
14. Lim B, Son S, Kim H, et al. Enhanced deep residual networks interpolation. Visual Computing for Industry, Biomedicine,
for single image super-resolution. In: Proceedings of the IEEE and Art. 2019: 2: 13.
Conference on Computer Vision and Pattern Recognition 25. Zeng GL. Directly filtering the sparse-view CT images by
Workshops, Honolulu, Hawaii. 2017; 136-144. BM3D. SL Clinical Medicine. 2022; 5: 126.

© 2023 Christiansen C, et al. This article is distributed under the terms of the Creative Commons Attribution 4.0 International License

J Biotechnology App, 2023 Volume 2 | Issue 1 | 7 of 7