A Survey on Image Registration using Machine Learning Algorithms

Image registration is the process of transforming different sets of data into one coordinate system. Data may be multiple photographs, data from different sensors, times, depths, or viewpoints. Image registration or image alignment algorithms can be classified into intensity-based and feature-based. In linear registration, it is generally difficult to optimize the large deformations between images with large anatomical differences. This is further complicated if images of two different imaging modalities with large appearance differences need to be registered. Second, with the advent of cutting-edge imaging technology, the scale and diversity of imaging data has increased significantly, posing significant challenges to registration. Modern registration methods need to be sufficiently versatile to deal with diverse imaging data with high efficiency, accuracy, and robustness. Fortunately, machine learning techniques applied to image registration tasks can help address the aforementioned issues. Specifically, different machine learning techniques can be employed to learn from prior registration results to improve the registration performance in some challenging tasks. This paper will be dedicated to summarize state-of-the-art learning-based Regis- ration algorithms, particularly methods for deformable registration. Methods for linear registration will also be discussed where necessary