Work

Image guided respiratory motion analysis: time series and image registration

Year: 2008
Type: dissertation
Abstract: The efficacy of Image guided radiation therapy (IGRT) systems relies on accurately extracting, modeling and predicting tumor movement with imaging techniques. This thesis investigates two key problems associated with such systems: motion modeling and image processing. For thoracic and upper abdominal tumors, respiratory motion is the dominant factor for tumor movement. We have studied several special structured time series analysis techniques to incorporate the semi-periodicity characteristics of respiratory motion. The proposed methods are robust towards large variations among fractions and populations; the algorithms perform stably in the presence of sparse radiographic observations with noise. We have proposed a subspace projection method to quantitatively evaluate the semi-periodicity of a given observation trace; a nonparametric local regression approach for real-time prediction of respiratory motion; a state augmentation scheme to model hysteresis; and an ellipse tracking algorithm to estimate the trend of respiratory motion in real time. For image processing, we have focused on designing regularizations to account for prior information in image registration problems. We investigated a penalty function design that accommodates tissue-type-dependent elasticity information. We studied a class of discontinuity preserving regularizers that yield smooth deformation estimates in most regions, yet allow discontinuities supported by data. We have further proposed a discriminate regularizer that preserves shear discontinuity, but discourages folding or vacuum generating flows. In addition, we have initiated a preliminary principled study on the fundamental performance limit of image registration problems. We proposed a statistical generative model to account for noise effect in both source and target images, and investigated the approximate performance of the maximum-likelihood estimator corresponding to the generative model and the commonly adopted M-estimator. A simple example suggests that the approximation is reasonably accurate. Our studies in both time series analysis and image registration constitute essential building-blocks for clinical applications such as adaptive treatment. Besides their theoretical interests, it is our sincere hope that with further justifications, the proposed techniques would realize its clinical value, and improve the quality of life for patients.
Language: English
Subfield: Radiation
Open Access status: green
OpenAlex ID: W139356225
Cites: 136
Cited by: 3
Related to: 10