PROJECT TITLE :

Learning Deformable Image Registration from Optimization Perspective, Modules, Bilevel Training and Beyond

ABSTRACT:

The goal of conventional deformable registration methods is to solve an optimization model that has been carefully designed on image pairs, and the computational costs associated with these methods are extremely high. On the other hand, more recent methods based on deep learning are able to provide quick deformation estimation. These heuristic network architectures are fully data-driven, and as a result, they lack explicit geometric constraints, which are required in order to generate plausible deformations (such as topology-preserving ones). In addition, learning-based approaches typically treat hyper-parameter learning as a black-box problem and call for a significant amount of computational and human effort to carry out a large number of training runs. We propose a new learning-based framework with the goal of optimizing a diffeomorphic model through multi-scale propagation in order to address the issues that have been presented thus far. To be more specific, we formulate diffeomorphic registration by introducing a generic optimization model and develop a series of learnable architectures to obtain propagative updating in the coarse-to-fine feature space. All of these contributions are presented in this paper. In addition to this, we suggest a new bilevel self-tuned training strategy that makes it possible to search for task-specific hyper-parameters in an effective manner. This training strategy improves the adaptability to different kinds of data while simultaneously reducing the amount of computational and human labor that is required. Experiments involving image registration are carried out on 3D volume datasets by our team in two distinct groups: image-to-atlas registration on brain MRI data and image-to-image registration on liver CT data. Both of these types of registration are carried out separately. Extensive results show that the proposed method performs at the cutting edge of its field, offering a diffeomorphic guarantee while maintaining an extremely high level of efficiency. We also apply our framework to difficult multi-modal image registration, and we investigate how our registration can support the downstream tasks for medical image analysis, such as multi-modal fusion and image segmentation. This is done by applying our framework to challenging multi-modal image registration.