Date of Award
Doctor of Philosophy (PhD)
X-ray computed tomography (CT) is an important and effective tool in medical and industrial
imaging applications. The state-of-the-art methods to reconstruct CT images have had
great development but also face challenges. This dissertation derives novel algorithms to
reduce bias and metal artifacts in a wide variety of imaging modalities and increase performance
in low-dose scenarios.
The most widely available CT systems still use the single-energy CT (SECT), which is
good at showing the anatomic structure of the patient body. However, in SECT image
reconstruction, energy-related information is lost. In applications like radiation treatment
planning and dose prediction, accurate energy-related information is needed. Spectral CT
has shown the potential to extract energy-related information.
Dual-energy CT (DECT) is the first successful implementation of spectral CT. By using two
different spectra, the energy-related information can be exported by reconstructing basis-material
images. A sinogram-based decomposition method has shown good performance in
clinical applications. However, when the x-ray dose level is low, the sinogram-based decomposition
methods generate biased estimates. The bias increases rapidly when the dose level
decreases. The bias comes from the ill-posed statistical model in the sinogram-decomposition
method. To eliminate the bias in low-dose cases, a joint statistical image reconstruction
(JSIR) method using the dual-energy alternating minimization (DEAM) algorithm is proposed.
By correcting the ill-posed statistical model, a relative error as high as 15% in the
sinogram-based decomposition method can be reduced to less than 1% with DEAM, which
is an approximately unbiased estimation.
Photon counting CT (PCCT) is an emerging CT technique that also can resolve the energy
information. By using photon-counting detectors (PCD), PCCT keeps track of the energy
of every photon received. Though PCDs have an entirely different physical performance
from the energy-integrating detectors used in DECT, the problem of biased estimation with
the sinogram-decomposition method remains. Based on DEAM, a multi-energy alternating
minimization (MEAM) algorithm for PCCT is proposed. In the simulation experiments,
MEAM can effectively reduce bias by more than 90%.
Metal artifacts have been a concern since x-ray CT came into medical imaging. When there
exist dense or metal materials in the scanned object, the image quality may suffer severe
artifacts. The auxiliary sinogram alternating minimization (ASAM) algorithm is proposed
to take advantages of two major categories of methods to deal with metal artifacts: the
pre-processing method and statistical image reconstruction. With a phantom experiment, it
has been shown that ASAM has better metal-artifact reduction performance compared with
the current methods.
A significant challenge in security imaging is that due to the large geometry and power
consumption, low photon statistics are detected. The detected photons suffer high noise and
heavy artifacts. Image-domain regularized iterative reconstruction algorithms can reduce
the noise but also result in biased reconstruction. A wavelet-domain penalty is introduced
which does not bring in bias and can effectively eliminate steaking artifacts. By combining
the image-domain and wavelet-domain penalty, the image quality can be further improved.
When the wavelet penalty is used, a concern is that no empirical way, like in the image-domain
penalty, is available to determine the penalty weight. Laplace variational automatic
relevance determination (Lap-VARD) method is proposed to reconstruct the image and
optimal penalty weight choice at the same time.
Joseph A. O'Sullivan
David G. Politte, Matthew Lew, Ulugbek Kamilov, Umberto Villa,