UCLA

As a fundamental task in medical image analysis, deformable image registration (DIR) is the process of estimating the deformation vector fields (DVFs) to images. In classic optimization-based DIR method, DVF is solved by optimizing a cost function consisting of image dissimilarity and DVF regularity, which typically involves time-consuming iterative processes. Deep-learning (DL)-based DIR has been developed in recent years, which offers a much faster alternative and the benefit from data-driven regularizing behaviors. This dissertation aims to develop accurate and robust DIR methods and address the lingering challenges in DL-DIR. First, we propose a DIR network that is conscious of and self-adaptive to deformation of various scales to improve accuracy. Second, we propose supervised and unsupervised approaches to incorporate learned implicit feasibility prior into DIR. Third, we propose a domain adaptation method to address the potential domain shift in DIR and improve accuracy and robustness on new data. Finally, we propose a DIR approach to synthesize continuous 4D motion from 3D image pair. Experiments with lung and cardiac images showed that the proposed techniques yielded significant performance improvement. We demonstrate the strength of combining physical-driven rationales and DL techniques in DIR.