Deep learning approach to discontinuity-preserving image registration
Image registration is an indispensable tool in medical image analysis. Traditionally, registration algorithms are aimed at aligning image pairs using regularizers to impose smoothness restrictions on unknown deformation fields. The majority of these methods assume global smoothness in the image domain, which pose issues for scenarios where motion discontinuities exist. Examples where local motion discontinuities happen are the sliding motion between adjacent organ tissues, and the pushing motion of the lungs against the chest wall during the respiratory cycle. Furthermore, an objective function must be optimized for each given pair of images. Thus registration of multiple image sets becomes very time-consuming and poorly scale with higher resolution image volumes. Using recent developments in deep learning, we propose an unsupervised learning-based image registration model. This model is trained over a loss function with a custom regularizer that preserves local discontinuities while simultaneously respecting the smoothness assumption in homogeneous regions of image volumes. In following a learning-based model, the image registration process can be completed within seconds, which is significantly quicker than optimization-based registration algorithms. The proposed model will be evaluated qualitatively and quantitatively on datasets of chest computed tomography (CT) 3D image volumes.