Dept. of Computer Science and Engineering
Oregon Health & Science University
Images provide vital information about this world. Multiple images often share the same scene observed at different times, from different view angles or using different sensors. Image registration is a method of aligning two or more images into the same coordinate system, so that the aligned images can be directly compared, combined and analyzed. Correspondence identication between the images is usually a simple task for human visual system, but for the computer algorithm it represents a challenging problem. Automated estimation of the correspondences between the imaged objects and recovery of the underlying geometrical transformation is a fundamental goal of image registration. In medical imaging, images are often related through complex non-rigid deformations. Method to recover such non-rigid geometrical transformations are called non-rigid image registration methods. In this thesis, we have developed several contributions to the field of non-rigid image registration. These contributions are linked under the common theme of non-rigid image registration, but stand on their own as valuable components within image registration framework. We have developed a new intensity-based similarity measure, called Residual Complexity (RC), to cope with images corrupted by spatially-varying intensity distortions. Such distortions are common in microscopy and magnetic resonance imaging, and represent many challenges for image registration. RC is optimized when the residual image can be sparsly coded using a few known basis functions, which explicitly account for spatially varying distortions. We have also developed a novel method for rigid and non-rigid point set registration, called Coherent Point Drift (CPD) algorithm. The algorithm simultaneously recovers the correspondences between two sets of multidimensional points as well as the underlying non-rigid transformation. CPD can be used as a key component in feature-based non-rigid image registration, but also has many applications in different computer vision areas. Finally, we have developed an automated system for motion estimation from 3D+T echocardiography. The system is based on sequential non-rigid image registration, and includes several new contribution, such as ultrasound-specific similarity measure, shape and dynamic constraints. The system outputs the dense deformation field, which we use to derive myocardium quantitative characteristics, such as strain and torsion. We have validated the accuracy of our approach with the groundtruth measurements from implanted markers.
School of Medicine
Myronenko, Andriy, "Non-rigid image registration regularization, algorithms and applications." (2010). Scholar Archive. 370.