Biomechanically-Regularized Deformable Image Registration for Head and Neck Adaptive Radiation Therapy

Abstract

Radiation treatment (RT), one of the best treatments available for head and neck (HN) cancer, may fail to accurately target tumors and spare surrounding healthy tissue that change in shape and location during 5-7 weeks of RT. This anatomical change can be monitored by calculating deformation maps from planning computed tomography (CT) image (taken prior to the start of RT) to treatment CT images (taken at every treatment fractions for patient setup) via deformable image registration (DIR). In response to the deformations estimated by DIR, initial radiation treatment plan established on the planning CT can be adjusted to deliver sufficient radiation dose to the tumors while sparing healthy tissue. However, since DIR is formulated as an optimization problem to find a deformation map that simply maximizes a similarity metric between two images, it may result in physically unreasonable deformations, such as bone warping. Moreover, DIR accuracy of HN soft tissue region is limited and parameter-dependent as reported in previous studies. Finally, previous studies have evaluated DIR accuracy with a limited number of landmarks, with which accuracy of volumetric deformation cannot be rigorously evaluated. The objective of this dissertation is 1) to improve registration accuracy of HN CT images by introducing penalty terms (from biomechanical principles) into B-spline DIR, in which deformation is represented using a linear combinations of B-spline functions, and 2) to develop an improved evaluation method for DIR accuracy based on finite element model (FE) model of HN region. First, a penalty for prevent the bone warping was developed to preserve inter-voxel distances within each of rigid regions. Second, a penalty that prevents resultant deformations from violating the static equilibrium equations of linear elastic material was used for the B-spline DIR of muscle in HN region. Third, a FE HN model was developed to generate deformation maps similar to those seen in patients that can be used as ground-truth for the evaluation of registration accuracy. The outcome of the dissertation would support research/development in RT of HN cancer by enabling the accurate estimation of deformations of healthy tissue surrounding tumor and the rigorous assessment of registration accuracy.