A Survey of Unsupervised Learning in Medical Image Registration

Introduction

Image registration is a technology that transforms different images into the same coordinate system according to the same content matching (Thomas et al. 2003). That is, the premise that images can be registered is that the content represented by different images is the same. In the field of medical image registration, different images may be taken by a certain body part using the same medical imaging equipment at different times, or a certain body part may be taken by different medical imaging equipment at different times. The difference between the images caused by the former may be the difference in the small spatial position changes of the corresponding parts, such as the annual physical examination, or the corresponding parts have undergone physiological and anatomical changes at different times, such as changes in the morphology of the lungs due to lung breathing (McInerney et al. 2016). The difference between the images caused by the latter is due to the difference in the imaging principles of the equipment, such as X-ray imaging, MRI and ultrasound imaging (Thomas 2004), etc., which cause the images to be differentiated into different modalities. Medical images of inter-modality have great differences in image intensity and location information.

However, medical image registration plays a very important role in many clinical applications. Single-modality images can only provide limited lesion information, while fusion processing of multi-modality medical images can provide more comprehensive diagnostic information. The images of patients can also be fused with standard anatomical atlas to assist surgical treatment. The premise of fusion (James et al. 2014) is that medical images of different modalities need to be effectively registered with the anatomical structure in space.

Actually, image registration is to find a suitable transformation relationship between different images. The traditional image registration method aligns different images by position, and then optimizes each voxel pair via an iterative method. Although this method can achieve good registration accuracy, its shortcomings are also very obvious. First, the process of iterative optimization makes the entire registration process too slow. Usually, the registration takes several hours. Such registration efficiency cannot be used on a large scale in clinical applications. Second, the registration is an independent rather than a learning process, which means that each registration process takes the same amount of time. The emergence of deep networks has changed the traditional registration mode, and transformation parameters can be learned in the network. The traditional registration process (Maes et al. 1997) of several hours is compressed to a few seconds. In the supervised registration method, the training data sent to the registration network has ground truth, that is, the registered images after spatial transformation. Usually, the approach adopted by supervised learning is to obtain ground truth of the training data in advance by using traditional iterative-based registration methods. However, the problem with supervised learning is that the ground truth itself is not a real label, but a transformation result with higher accuracy, so the effect of registration will be greatly reduced due to the inaccuracy of the label. This leads to another registration method explored in this article, the unsupervised registration method (Geert et al. 2017; Grant et al. 2019; Nicholas et al. 2019). The unsupervised method is an improved version based on the defects of the supervised method. The unsupervised method uses the difference between different images to construct a deformation field, where the deformation field is a transform vector field. The common network architecture of the deformation field will be given in the next section. The third section of this article introduces various methods of registering objects with different relationships.