Regression learning for 2D/3D image registration.

摘要

Image registration is a common technique in medical image analysis. The goal of image registration is to discover the underlying geometric transformation of target objects or regions appearing in two images. This dissertation investigates image registration methods for lung Image-Guided Radiation Therapy (IGRT). The goal of lung IGRT is to lay the radiation beam on the ever-changing tumor centroid but avoid organs at risk under the patient's continuous respiratory motion during the therapeutic procedure.;To achieve this goal, I developed regression learning methods that compute the patient's 3D deformation between a treatment-time acquired x-ray image and a treatment-planning CT image (2D/3D image registration) in real-time. The real-time computation involves learning x-ray to 3D deformation regressions from a simulated patient-specific training set that captures credible deformation variations obtained from the patient's Respiratory-Correlated CT (RCCT) images. At treatment time, the learned regressions can be applied efficiently to the acquired x-ray image to yield an estimation of the patient's 3D deformation. In this dissertation, three regression learning methods - linear, non-linear, and locally-linear regression learning methods are presented to approach this 2D/3D image registration problem.