Author's School

School of Engineering & Applied Science

Author's Department/Program

Electrical and Systems Engineering

Language

English (en)

Date of Award

Summer 9-1-2014

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Chair and Committee

Joseph A O'Sullivan

Abstract

This dissertation contributes toward solutions to two distinct problems linked through the use of common information optimization methods. The first problem is the X-ray computed tomography (CT) imaging problem and the second is the computation of Berger-Tung bounds

for the lossy distributed source coding problem. The first problem discussed through most of the dissertation is motivated by applications in radiation oncology, including dose prediction in proton therapy and brachytherapy.

In proton therapy dose prediction, the stopping power calculation is based on estimates of the electron density and mean excitation energy. In turn, the estimates of the linear attenuation coefficients or the component images from dual-energy CT image reconstruction are used

to estimate the electron density and mean excitation. Therefore, the quantitative accuracy of the estimates of the linear attenuation coefficients or the component images affects the accuracy of proton therapy dose prediction.

In brachytherapy, photons with low energies (approximately 20 keV) are often used for internal treatment. Those photons are attenuated through their interactions with tissues. The dose distribution in the tissue obeys an exponential decay with the linear attenuation coefficient as the parameter in the exponential. Therefore, the accuracy of the estimates of

the linear attenuation coefficients at low energy levels has strong influence on dose prediction in brachytherapy.

Numerical studies of the regularized alternating minimization (DE-AM) algorithm with different regularization parameters were performed to find ranges of the parameters that can achieve the desired image quality in terms of estimation accuracy and image smoothness.

The DE-AM algorithm is an extension of the AM algorithm proposed by O'Sullivan and Benac. Both simulated data and real data reconstructions, as well as system bias and variance experiments, were carried out to demonstrate that the DE-AM algorithm is

incapable of reconstructing a high density material accurately with a limited number of iterations (1000 iterations with 33 ordered subsets). This slow convergence phenomenon was then studied via a toy. or scaled-down problem, indicating a highly ridged objective function.

Motivated by the studies which demonstrate the slow convergence of the DE-AM algorithm, a new algorithm, the linear integral alternating minimization (LIAM) algorithm was developed, which estimates the linear integrals of the component images first; then the component

images can be recovered by an expectation-maximization (EM) algorithm or linear regression methods. Both simulated and real data were reconstructed by the LIAM algorithm while varying the regularization parameters to ascertain good choices ( &delta= 500; &lambda= 50 for I0 =

100000 scenario). The results from the DE-AM algorithm applied to the same data were used for comparison. While using only 1/10 of the computation time of the DE-AM algorithm, the LIAM algorithm achieves at least a two-fold improvement in the relative absolute error

of the component images in the presence of Poisson noise.

This work also explored the reconstruction of image differences from tomographic Poisson data. An alternating minimization algorithm was developed and monotonic decrease in the objective function was achieved for each iteration. Simulations with random images

and tomographic data were presented to demonstrate that the algorithm can recover the difference images with 100% accuracy in the number of and identity of pixels which differ. An extension to 4D CT with simulated tomographic data was also presented and an approach

to 4D PET was described.

Different approaches for X-ray adaptive sensing were also proposed and reconstructions of simulated data were computed to test these approaches. Early simulation results show improved image reconstruction performance in terms of normalized L2 norm error compared to a non-adaptive sensing method.

For the second problem, an optimization and computational approach was described for characterizing the inner and outer bounds for the achievable rate regions for distributed source coding, known as Berger-Tung inner and outer bounds. Several two-variable examples

were presented to demonstrate the computational capability of the algorithm. For each problem considered that has a sum of distortions on the encoded variables, the inner and outer bound regions coincided. For a problem defined by Wagner and Anantharam with

a single joint distortion for the two variables, their gap was observed in our results. These boundary regions can motivate hypothesized optimal distributions which can be tested in the first order necessary conditions for the optimal distributions.

Comments

Permanent URL: http://dx.doi.org/10.7936/K7JH3J8Q

Share

COinS