Files
Abstract
Image registration, the task of establishing spatial correspondences between two images, is one of the most fundamental research problem in medical image analysis. Diffeomorphic image registration (DiffIR) guarantees the preservation of topologies between the corresponding structures in medical images and has a large number of clinical applications. However, traditional DiffIR methods are time-consuming and impractical because of their complex optimization in the high-dimensional image space. Recently, deep learning (DL) based approaches have been proposed that greatly improve the computational efficiency of DiffIR. However, there are three main challenges faced by existing DL-based DiffIR. Firstly, they are rooted on fluid-based DiffIR but have limited flexibility for capturing large deformations because of its assumption on deformations driven by stationary velocity fields (SVF). Secondly, they are not efficient enough to handle high-resolution 3D image volumes and often sacrifice registration accuracy by halving DL model, image, or deformation size to reduce the memory cost. Thirdly, most existing methods focus on registering images without missing correspondences and have difficulty handling images with appearing or disappearing objects or structures, such as those in the presence of pathologies, i.e., metamorphic image registration. This dissertation addresses the above three challenges in a unified DL framework, which provides flexible and efficient diffeomorphic registration solutions for high-resolution images on a large scale, with and without missing correspondences. First, to increase the flexibility of DL-based DiffIR, this dissertation adopts residual network blocks with Lipschitz continuity (LC-ResNet blocks) to solve the ordinary differential equations that govern the dynamics of deformations driven by velocity fields. Based on LC-ResNet Blocks, a residual registration network (R2Net) is proposed with two variants, SVF-R2Net and NSVF-R2Net, to model flexible parameterizations of the deformations, either based on stationary or time-varying (non-stationary) velocity fields. The effectiveness of the proposed networks are evaluated on a wide range of anatomies and modalities of images, including cardiac ACDC MRI dataset, brain OASIS MRI and lung EMPIRE10 CT images with improved registration accuracy and smoother deformations compared to existing SOTA methods.
Second, to provide an efficient solution for high-resolution images, this dissertation proposes a Diving and Downsampling mixed Registration Network (DDR-Net), which separates the velocity estimation and deformation integration into two steps by using a two-phase learning design. The first phase takes the registration tasks of the chunked and downsampled images with smaller size, while the second phase takes the velocity fields output from phase one and integrates them back to the original size. The results of improved accuracy and deformation smoothness are shown on the OASIS 3D brain MRI dataset. The combination of R2Net and DDR-Net results in two multi-scale variants, MS-SVF-R2Net and MS-NSVF-R2Net, which are evaluated on the IBSR18 brain MRI dataset with improved registration accuracy and reduced time and memory cost.
Third, to handle image metamorphosis with missing correspondences like the tumor regions, this dissertation provides a metamorphic image registration network (MetaRegNet) for aligning a pair of a healthy source image and a target image with pathology. This network is built upon R2Net, but the difference is, it jointly learns the smooth spatial transformations between an image pair and their changes or differences in the intensity variations. The successful application of this model has been shown on the BraTS2021 Brain Tumor MRI dataset, with smoother estimation of deformations, especially around the pathological regions, and better appearance estimation for pathology.