Weighted Least Squares Kinetic Upwind Method using Eigen Vector Basis
Generalized Penalized Weighted Least-Squares...
Transcript of Generalized Penalized Weighted Least-Squares...
Generalized Penalized Weighted Least-Squares
Reconstruction for Deblurred Flat-Panel CBCT Steven Tilley II, Jeffrey H. Siewerdsen, J. Webster Stayman
Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD
Abstract—An increase in achievable spatial resolution would
enable application of flat-panel detector cone-beam CT (FP-
CBCT) in high-resolution applications, such as 3D breast and bone
imaging. Improved reconstruction techniques exploiting more
sophisticated models of noise and spatial resolution characteristics
may provide such improvement. In this work, we propose a
statistical model of measurement mean and covariance, with
particular attention given to correlations in measurement data. An
iterative reconstruction algorithm based on this model was applied
to simulated and real experimental data to evaluate its
performance compared to an uncorrelated model under various
imaging conditions. Focal spot and detector blur of a CBCT
testbench were measured and used to inform the model-based
image reconstruction. Simulation studies demonstrate that
incorporation of the correlated noise model yields improved
reconstructions, especially in the presence of focal spot blur.
Improved noise-resolution tradeoffs were quantified, and
application to high-resolution imaging of trabecular bone was
demonstrated.
Keywords— Spatial resolution, Cone-beam CT, Correlated noise,
Model-based Reconstruction, Generalized Least-Squares Estimation
I. INTRODUCTION
Flat-panel cone-beam computed tomography (FP-CBCT)
has found many applications in high spatial resolution imaging
due to the large volumetric coverage and fine detector pixels.
Despite this advantage, the spatial resolution required for some
applications remains just out of reach. For example,
visualization of microcalcifications in CBCT mammography or
highly detailed trabecular structure in quantitative CT of
musculoskeletal extremities both stand to benefit from
improvements beyond conventional spatial resolution limits.
Current breast imaging solutions include projection
mammography, tomosythesis, and CBCT, with the requirement
to detect microcalcifications at a scale of ~0.1 mm. Similarly in
trabecular bone imaging, the gold standard is microCT which
tends to have a very small field of view, making it difficult to
apply to in vivo applications. In this paper, we seek to extend
the high spatial resolution capabilities of FP-CBCT through
careful modeling of system blur and correlated noise within a
statistical reconstruction framework.
Typical model-based iterative reconstruction algorithms
balance accuracy and complexity of the imaging system model
based on desired resolution, noise properties, algorithm
complexity, and execution time. The standard approach is to
adopt a projector model with little or no blur and independent
measurement noise. Such assumptions have been supported in
part by work done by Hofmann et al. [1], which shows that
modeling source blur in clinical CT systems does not
significantly improve image quality for typical focal spot sizes
(~0.5 mm) and detector pixels sizes (~0.6 mm). While this may
be true for current diagnostic CT systems and current spatial
resolution goals, there are a number of factors that suggest an
advantage by way of improved system models for FP-CBCT.
Such factors include 1) finer detector pixels (70-200 m) in flat
panels; 2) larger measurement blur due to light spread in the
scintillator; 3) increased focal spot blur due to widespread use
of inexpensive fixed anode x-ray tubes and compact geometries
in dedicated FP-CBCT systems; and 4) the presence of
significant noise correlation in FP-CBCT data due to light
spread after the conversion of primary x-ray quanta to
secondary light photons.
Model-based reconstruction approaches that account for
system blur but without a correlated noise model have been
previously developed. [2], [3] Similarly, sinogram restoration
methods have also been proposed to correct for system blur [4]
and noise correlation [5]. We propose a combination of
sinogram restoration (deblurring) followed by iterative
reconstruction, tracking the correlations through the deblur
processing and accounting for them in the reconstruction step.
Flat-panel CBCT systems can be characterized by multiple
sources of blur and noise. We consider blur due to an extended
x-ray source and due to the detector scintillator, the latter of
which adds correlations to the noise. Because noise correlation
is a result of only one of these two sources of modeled blur, we
explore the effects of varying the magnitude of these blurs in
simulation. Understanding the scenarios that benefit from blur
and correlation models will not only improve reconstructions
on current FP-CBCT systems, but will help to guide new system
design (e.g., geometry, focal spot size, and scintillator
thickness).
In the following sections we present an overview of the
reconstruction methodology. This approach is a refined version
of the methods reported in [6] with a modified strategy for
deblurring. A simulation investigation for various system
designs with different magnitudes of source and detector blur
follows. Subsequently, a brief description of the physical
measurement of source and detector blur for a FP-CBCT
testbench is presented. Finally, the measured blur models are
incorporated within the proposed methodology and applied to
FP-CBCT extremity data for trabecular bone imaging.
II. METHODS
A. Imaging System Model and Reconstruction
We consider a system model based on the first and second
order statistics. Thus, the measurements are specified by the
mean vector (�̅�) and the covariance matrix (𝐊𝑌 ) which are
functions of the attenuation (), related by a forward projection
operator (A), a gain operator (G), and the source and detector
scintillator blurs (Bs and Bd respectively). These models can be
represented compactly in vector form as
�̅� = 𝐁𝑑𝐁𝑠𝐆 exp(−𝐀μ̅) (1)
𝐊𝑌 = 𝐁𝑑𝐷{𝐁𝑠𝐆 exp(−𝐀μ̅)}𝐁𝑑𝑇 + 𝐊𝑟𝑜 (2)
where Kro denotes additive readout noise in the detector. (Note
that source blur is approximated as a projection domain effect –
implying that depth dependent magnification effects are small.)
Transforming measurements into line integral estimates and
presuming Gaussian noise permits accommodation of a non-
diagonal covariance matrix within a linear least-squares
framework. However, such linearization requires an initial
deblur operation. We introduce a “modified” deblur operator
𝐂−1 ≈ 𝐁𝑑−1𝐁𝑠
−1 (3)
which is approximately equal to the inverse of the source and
detector blurs. Exact inversion may not be possible or desirable
due to null spaces and noise magnification.
The line integral estimate and its covariance then become
𝑙 = − log(𝐆−1𝐂−1𝑦) (4)
𝐊𝐿 ≈ 𝐷 {
1
𝐂−1𝑦} 𝐂−1𝐊𝑌[𝐂−1]𝑇𝐷{
1
𝐂−1𝑦}
(5)
yielding the penalized weighted least-squares equation:
�̂� = arg min𝜇
||𝑙 − 𝐀𝜇||𝐊𝐿
−1+ 𝛽𝑅(𝜇) (6)
B. Practical Implementation
In this work, we use a quadratic penalty function, allowing
equation (6) to be written in a closed form.
�̂� = [𝐀𝑇𝐊𝐿−1𝐀 + 𝛽𝐑]−1𝐀𝑇𝐊𝐿
−1𝑙 (7)
We assume all blurs are shift invariant, allowing blur operations
to be done in the Fourier domain. We choose the deblur
operation in (3) so that it applies a binary mask (M) zeroing any
nulled frequencies, approximated as the frequencies where the
blur MTF is less than 10-2 (simulation) or 10-4 (testbench data).
𝑪−1𝑥 = 𝐹−1 (𝑀𝐹(𝑥)
𝐹(𝑩𝑠𝑩𝑑))
(8)
where F denotes the Fourier transform. Consequently:
[𝑪−1]−1 ≈ 𝑪∗𝑥 = 𝐹−1(𝐹(𝑩𝑠𝑩𝑑)𝑀 𝐹(𝑥)) (9)
which allows the inverse covariance matrix to be written as
𝐊𝐿−1 ≈ 𝐷{𝐂−1𝑦}[𝐂∗]𝑇 𝐊𝑌
−1[𝐂∗]𝐷{𝐂−1𝑦} (10)
where KY is approximated as
𝐊𝑌 ≈ 𝐁𝑑𝐷{𝐂−1𝑦}𝐁𝑑𝑇 + 𝐊𝑟𝑜 (11)
The null space in C* forbids an inversion of 𝐊𝐿 , but using
iterative methods we may compute a generalized inverse and
solve the likelihood equation within the unmasked subspace.[7]
A similar approach was used for null spaces in Bd when
inverting 𝐊𝑌 in simulation studies.
Solving of (7) can be divided into two parts, preprocessing
and iterative steps. In the preprocessing step line integral
estimates are computed from the measurements using the
modified deblur, the inverse covariance matrix is applied, and
the result is backprojected. The inverse in (7) is solved
iteratively using the conjugate gradient method (100 iterations
for simulated data, 300 for real data). Each iteration requires a
convolution with the penalty kernel, a forward projection, an
application of the inverse covariance, and a backprojection. The
application of the covariance matrix requires an application of
the inverse of 𝐊𝑌, also via the conjugate gradient method (1000
iterations in the preprocessing step and 100 in each iteration of
the iterative step, or until the residual decreased to 0).
C. Simulation Studies
Figure 1 shows the simulation phantom. Projection data were
generated using a high resolution version of the phantom
(25 m voxels) onto a high resolution detector (1x7000 line of
35 m pixels at 720 angles), which were then downsampled to
1x1750 at 720 angles. The data were reconstructed into a
1000x1000 grid of 0.1 mm voxels. Gaussian noise was added
to the data to simulate correlated quantum noise and white
readout noise. Noiseless data generation from line integrals is
shown in (12), and noise injection is show in (13).
�̅� = 𝐁𝑑𝐁𝑠𝐆exp (−𝑙) (12)
𝑦 = �̅� + 𝑩𝑑𝑁 (0, √𝐁𝑠𝐆 exp(−𝑙)) + 𝑁(0, 𝜎𝑟𝑜) (13)
The variance of the readout noise was 3.5 photons2. Gain was a
constant 106 photons.
Two reconstruction methods were studied: 1) using the full
covariance matrix in equation (10); and 2) an uncorrelated
covariance matrix composed of only diagonal elements:
𝐊𝐿−1 ≔ 𝐷{𝐂−1𝑦} + 𝐷{𝜎𝑟𝑜
2 } (14)
which represents an independent noise assumption. When this
noise model is used, (7) is a traditional penalized least-squares
reconstruction with deblurred measurements. The noise models
were compared by plotting two resolution-variance curves from
reconstructions with varying penalty strength.
Resolution was measured as an edge response to the disc in
the phantom, and variance was approximated as the spatial
variance of the data inside the blue ring shown in Figure 1.
Additional FDK reconstructions were conducted with the cutoff
frequency at Nyquist and no apodization.
To evaluate the importance of the two types of blurs, various
blur scenarios were tested. There were two subsets of
experiments where: 1) the total blur was kept constant, and the
distribution of blur was altered from source dominated to
detector dominated; and 2) the ratio of the blurs was kept
constant, and the value of the total blur was varied. All blurs in
the simulation study were Gaussian.
D. Testbench Characterization
In order to apply this algorithm to physical FP-CBCT data,
we estimated the source and detector blurs on our x-ray
testbench. This was done by imaging a tungsten edge,
calculating the edge spread function (ESF), and then converting
this to a modulation transfer function (MTF).[8] The scintillator
Figure 1: Simulation
phantom with fat (a),
muscle (b),
trabecular bone (c),
line pairs (d), and a
uniform disc (e) for
resolution (edge).
and noise
measurements
(a)
(c)
(d) (e)
(b)
blur was assumed to be radially symmetric. The source blur is
known to not be radially symmetric, so MTFs were taken in the
(approximate) vertical, horizontal, and 45 degree directions.
These results were compared with MTFs taken from a pinhole
image of the source, acquired by placing a lead plate with a
pinhole on the filter and using a geometry with a high
magnification.
E. Physical FP-CBCT testbench data studies
Our testbench includes a Varian 4030CB detector and a
Varian Rad-94 x-ray tube. A custom wrist phantom (The
Phantom Laboratory, Greenwich, NY) was imaged with large
(~1.1 mm) and small (~0.5 mm) focal spots. We used the large
focal spot for testing our method, and the small focal spot to
create a high resolution reference image using FDK. The data
were reconstructed into a 210x600x600 volume of .15 mm
voxels using the correlated and uncorrelated noise models with
readout noise modeled as a constant 3.5 photons2. These two
iterative reconstruction methods were noise matched.
Deblurred and non-deblurred data were also reconstructed
using FDK. FDK reconstructions were done using a cutoff at
Nyquist and no apodization.
III. RESULTS
Figure 2 shows resolution variance curves for simulated
systems with different combinations of source and detector
blur. Points positioned on the curve in the diagram all have the
same total blur, while points on the straight line have the same
ratio of source to detector blur. The correlated noise model
equals or outperforms the uncorrelated noise model in all cases.
When source blur is large there is a significant advantage to the
correlated noise model whereas when source blur is minimal,
the two methods are nearly equivalent. This is consistent with
the fact that only the detector blur adds correlation and when
source blur is absent, the deblurring operation whitens the
noise, making the uncorrelated noise assumption accurate.
When detector blur is absent, the data is uncorrelated and
deblurring source blur adds off-diagonal values to the
covariance matrix making the correlated noise model important.
When the ratio of the blurs is kept constant and total blur
Figure 4: Testbench MTF Measurements. A: Detector MTF, with gauss*sinc model. B: Source pinhole image. C: Source MTF from pinhole
image. D: Profiles from pinhole MTF. E: Source MTF from tungsten edge response.
Figure 3: Simulated data reconstructions.
Top half is difference with truth. Bottom
half is reconstruction with a zoomed
image of the line pair target. Values are in
unites of mm-1. The uncorrelated model
yields higher noise in resolution matched
reconstructions and coarser resolution in
noise matched reconstructions as
compared with the correlated noise model
images.
Figure 2: Resolution-Variance
plots. Subplots A-G correspond to
different systems with varying
source/detector blurs (top left).
Uncorrelated noise model:
triangles, correlated noise model:
squares, FDK: star. Note that
everything to left of FDK
represents an improved spatial
resolution and that the correlated
noise model has the greatest
advantage in the case of systems
dominated by source blur (E).
Smaller total blur (F) permits a
finer resolution limit while larger
blur (G) increases this fine limit.
increases, the relative performance of the methods is similar.
The fine resolution limit increases as blur increases.
Figure 3 shows reconstructions from a simulated system with
approximately twice as much source blur as detector blur (point
D in Figure 2). When the images are resolution matched, the
correlated noise model produces a reconstruction with less
noise. (See background in the difference images.) When noise
matched, using the correlated noise model produces a sharper
reconstruction (note line pairs) than the uncorrelated noise
model. Also note, while variances are matched the correlations
in the reconstruction differ substantially.
Figure 4 summarizes the testbench characterization studies.
Figure 4A shows the detector MTF (scintillator blur and pixel
sampling) measured using the tungsten edge. The solid line
shows a fit to Gaussian times a sinc function (representing the
detector aperture). The FWHM of this Gaussian (in the spatial
domain) is 0.85 pixels, or 0.33 mm. Figure 4E shows the source
MTF measured using the tungsten edge placed at the center of
rotation. The MTF is asymmetric. The 45 degree MTF has its
first zero at a higher frequency than the vertical or horizontal
MTFs, which is consistent with a source resembling a 2D rect
function. The pinhole image of the source (Figure 4B) supports
this. The “horns” at the top and bottom of the image may
explain the differences in the first side lobes of the vertical and
horizontal MTFs. The MTF was also calculated (although at a
different magnification) by taking the Fourier transform of the
pinhole image (Figure 4C). The vertical, horizontal, and 45
degree line profiles of this 2D MTF are shown in Figure 4D,
and approximately match the shape of the MTF calculated using
the edge spread. The solid line in Figure 4E is the Gaussian blur
used in the reconstruction of our wrist phantom. The blur
focuses on the main lobe and is slightly larger than reality,
which decreases instability due to inaccuracies in the model.
Figure 5 shows real data reconstructions. Figure 5A is a high
resolution reference image, with a zoom region (Figure 5B) and
patch for variance estimation identified. A coarser resolution
FDK reconstruction where most of the high resolution
trabecular structure has been lost is shown in Figure 5C.
Attempting to deblur (using the Gaussian approximations
discussed previously) prior to applying FDK results in
significant noise amplification (Figure 5D). In comparison,
noise matched uncorrelated (Figure 5E) and correlated (Figure
5F) noise model reconstructions are also shown. Again, the
correlated noise model achieves higher resolution - showing
finer structures than the uncorrelated case at matched noise.
Comparisons with the high resolution reference indicate that
these fine structures are representative of true trabecular details.
IV. DISCUSSION
In this work we have presented a general framework for
modeling both source and detector blur including correlated
noise effects. Investigations in both simulations and in physical
FP-CBCT data show that proper noise modeling can improve
the noise-resolution trade-off. This advantage is most dramatic
for systems with significant source blur, though deblurring and
an accurate noise model outperforms conventional FBP in all
cases considered.
This methodology has potential application in a variety of
applications requiring high spatial resolution, including
trabecular bone imaging and breast CBCT. Moreover, the
studies presented here begin to characterize the improvements
possible for specific system designs (e.g., varying amounts of
source and detector blur). This may have potential impact in
system design – for example, permitting use of larger focal spot
x-ray sources (improved power requirements) and thicker
scintillators (improved detection efficiency) – if the proposed
reconstruction methodology is integrated into the system.
V. ACKNOWLEDGEMENTS
This work was supported in part by NIH grants R21-EB-014964 and T32-EB010021, and an academic-industry partnership with Varian Medical Systems
(Palo Alto, CA.)
REFERENCES [1] C. Hofmann, M. Knaup, and M. Kachelriess, “Do We Need to Model the
Ray Profile in Iterative Clinical CT Image Reconstruction?,” 2013, p. 403.
[2] B. Feng, J. a Fessler, and M. a King, “Incorporation of system resolution
compensation (RC) in the ordered-subset transmission (OSTR) algorithm for transmission imaging in SPECT.,” IEEE Trans. Med. Imaging, vol.
25, no. 7, pp. 941–9, Jul. 2006. [3] D. F. Yu, J. a Fessler, and E. P. Ficaro, “Maximum-likelihood
transmission image reconstruction for overlapping transmission beams.,”
IEEE Trans. Med. Imaging, vol. 19, no. 11, pp. 1094–105, Nov. 2000. [4] P. J. La Rivière, J. Bian, and P. a Vargas, “Penalized-likelihood sinogram
restoration for computed tomography.,” IEEE Trans. Med. Imaging, vol.
25, no. 8, pp. 1022–36, Aug. 2006. [5] H. Zhang, L. Ouyang, J. Ma, J. Huang, W. Chen, and J. Wang, “Noise
correlation in CBCT projection data and its application for noise reduction
in low-dose CBCT.,” Med. Phys., vol. 41, no. 3, p. 031906, Mar. 2014. [6] S. Tilley II, J. H. Siewerdsen, and J. W. Stayman, “Iterative CT
Reconstruction using Models of Source and Detector Blur and Correlated
Noise.” [7] C. R. Rao, Linear Statistical Inference and Its Applications, 2nd ed. New
York: John Wiley & Sons, Inc., 1973, pp. 527–528.
[8] E. Samei, M. J. Flynn, and D. a. Reimann, “A method for measuring the presampled MTF of digital radiographic systems using an edge test
device,” Med. Phys., vol. 25, no. 1, p. 102, 1998.
(a) (b)
(d)
(c)
Figure 5: Testbench data reconstructions. In split images, top half
is difference with the high resolution reference image. Units are mm-1.
(a): soft tissue. (b) cartilage. (c) trabecular bone. (d) cortical bone.