Search results

(1 - 1 of 1)

Title

COMPUTER MODELING OF BREAST LESIONS AND STUDIES OF ANALYZER-BASED X-RAY IMAGING

Creator

Garcia, Luis De Sisternes

Date

2011-11, 2011-12

Description

Phase-contrast x-ray imaging is an emerging technique that promises to yield highly sensitive medical images of soft tissue, which is...
Show morePhase-contrast x-ray imaging is an emerging technique that promises to yield highly sensitive medical images of soft tissue, which is difficult to observe via conventional radiography given its low X-ray attenuation differences. One of these phase-contrast techniques, known as analyzer-based imaging, has demonstrated that highly detailed breast tissue images can be obtained using synchrotron radiation. However, synchrotron facilities are impractical for clinical use. This thesis introduces studies and exposure consideration towards the application of analyzer-based imaging in a clinical environment, particularly in the context of breast imaging. It also introduces a computational breast lesion model that generates randomized three-dimensional phantoms which follow realistically the characteristics observed in real lesions. Moving analyzer-based imaging to clinical application requires the consideration of photon noise, inherent from the use of a photon-limited conventional source. We summarize the statistical properties in the presence of photon noise of two popular analyzer-based imaging techniques, known as diffraction-enhanced imaging (DEI) and multiple-image radiography (MIR). The statistics for MIR have not been previously derived and are introduced in this thesis. Comparison of the resulting statistical predictions with results obtained by Monte Carlo simulation validated the analysis. An expression for the maximum-likelihood (ML) solution for analyzer-based imaging is presented as a way of minimizing the effects of photon noise in the reconstruction of the object’s absorption, refraction and ultra-small angle scattering properties, and more practical maximum-likelihood expectation-maximization (ML-EM) and maximum-a-posteriori expectation-maximization (MAP-EM) solutions are also introduced. The behavior of the ML-EM and MAP-EM solutions was compared to the results produced by the five best-known analyzer-based reconstruction methods using computer simulations. The ML-EM and MAP-EM reconstructions proved closer to the theoretical values as they do not rely on commonly known limitations and approximations introduced by the other techniques. We introduce the development and evaluation of a new computational breast lesion phantom model that can simulate either massess or microcalcifications. The proposed tool allows the generation of a large number of randomized three-dimensional breast lesion simulations following desired characteristics normally used to describe breast lesions in clinical practice. The initial motivation for the development of this new phantom model was to enable the proposed evaluations of analyzer-based imaging to be achieved. However, the model became a major focus of this thesis because it improves significantly upon those that can be found in previous literature. The proposed lesion model can be used for evaluation studies across different breast imaging techniques, as well as for training purposes, so it is our hope that it could become an important resource for the broader mammography research community. As part of the lesion modeling research, we also introduce methods to computationally modify experimental mammography and analyzer-based images of breast tissue so that they present the generated tumor simulations embedded within their parenchyma realistically. The realism of the simulated lesion images was evaluated by comparison of 83 real tumor cases observed in mammograms with 83 constructed hybrid images in which simulated tumors matching the characteristics observed in the real cases were embedded, with healthy tissue acting as background. As a quantitative comparison, extracted features describing tumor shape and density showed no statistically significant differences between real and simulated tumors. A known computational tumor classification technique based on their shape observed in mammography was implemented and showed no significant performance differences between real and simulated cases, as well as showing good correlation with previously published performance results in real tumors. To measure the realism for use in human observer studies, we conducted a reader study in which 5 experienced radiologists were asked to judge whether each of the 166 images was real or simulated by assigning a score on a 7-point scale. The results were analyzed in a multiple-reader multiple-case statistical framework. The conclusion of the study was that the readers’ accuracy in assessing whether the lesions were real or simulated was not significantly better than random chance. This thesis also incorporates a reader study to evaluate the degree to which photon-limited analyzer-based images may be effective for visualization of breast cancer features. Our motivation was to establish the x-ray intensity that would be required to make these methods feasible, the purpose being to serve as a guide in parameter selection for future design of imaging hardware. We conducted a series of observer studies that quantify the performance of analyzer-based refraction images at different noise levels for the task of identifying subtle details present in breast tumors which are relevant to clinical diagnosis. The cases shown to the readers consisted of hybrid images where simulated lesions of known characteristics were computationally embedded in real breast analyzer-based background images. The original phase-contrast data was obtained using synchrotron radiation and was later modified to simulate the noise and blurring effects produced from a photon-limited source with a 300μm aperture size, similar to those used in a laboratory environment. Results showed that the analyzer-based imaging techniques statistically outperformed conventional mammography for the given task with an average of just 128 recorded photons per pixel in background image regions
Ph.D. in Electrical Engineering, December 2011
Show less