Date on Master's Thesis/Doctoral Dissertation

5-2018

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Electrical and Computer Engineering

Degree Program

Electrical Engineering, PhD

Committee Chair

Amini, Amir

Committee Co-Chair (if applicable)

Zurada, Jacek

Committee Member

Zurada, Jacek

Committee Member

Inanc, Tamer

Committee Member

Frigui, Hichem

Committee Member

Dunlap, Neal

Committee Member

Wang, Brian

Author's Keywords

medical Imaging; deformable image registration; image segmentation; physiologic imaging including assessment of pulmonary mechanics; image processing

Abstract

This thesis is concerned with development of techniques for efficient computerized analysis of 4-D CT data. The goal is to have a highly automated approach to segmentation of the lung boundary and lung nodules inside the lung. The determination of exact lung tumor location over space and time by image segmentation is an essential step to track thoracic malignancies. Accurate image segmentation helps clinical experts examine the anatomy and structure and determine the disease progress. Since 4-D CT provides structural and anatomical information during tidal breathing, we use the same data to also measure mechanical properties related to deformation of the lung tissue including Jacobian and strain at high resolutions and as a function of time. Radiation Treatment of patients with lung cancer can benefit from knowledge of these measures of regional ventilation. Graph-cuts techniques have been popular for image segmentation since they are able to treat highly textured data via robust global optimization, avoiding local minima in graph based optimization. The graph-cuts methods have been used to extract globally optimal boundaries from images by s/t cut, with energy function based on model-specific visual cues, and useful topological constraints. The method makes N-dimensional globally optimal segmentation possible with good computational efficiency. Even though the graph-cuts method can extract objects where there is a clear intensity difference, segmentation of organs or tumors pose a challenge. For organ segmentation, many segmentation methods using a shape prior have been proposed. However, in the case of lung tumors, the shape varies from patient to patient, and with location. In this thesis, we use a shape prior for tumors through a training step and PCA analysis based on the Active Shape Model (ASM). The method has been tested on real patient data from the Brown Cancer Center at the University of Louisville. We performed temporal B-spline deformable registration of the 4-D CT data - this yielded 3-D deformation fields between successive respiratory phases from which measures of regional lung function were determined. During the respiratory cycle, the lung volume changes and five different lobes of the lung (two in the left and three in the right lung) show different deformation yielding different strain and Jacobian maps. In this thesis, we determine the regional lung mechanics in the Lagrangian frame of reference through different respiratory phases, for example, Phase10 to 20, Phase10 to 30, Phase10 to 40, and Phase10 to 50. Single photon emission computed tomography (SPECT) lung imaging using radioactive tracers with SPECT ventilation and SPECT perfusion imaging also provides functional information. As part of an IRB-approved study therefore, we registered the max-inhale CT volume to both VSPECT and QSPECT data sets using the Demon's non-rigid registration algorithm in patient subjects. Subsequently, statistical correlation between CT ventilation images (Jacobian and strain values), with both VSPECT and QSPECT was undertaken. Through statistical analysis with the Spearman's rank correlation coefficient, we found that Jacobian values have the highest correlation with both VSPECT and QSPECT.

Included in

Biomedical Commons

Share

COinS