Unsupervised Learning-Based Large Deformation Diffeomorphic Metric Mapping for Fast Image Registration

Diffeomorphic image registration is a fundamental tool in medical image processing that facilitates medical image analysis tasks as it preserves the image topology during the registration. In this work, we focused on the traditional large deformation diffeomorphic metric mapping (LDDMM) and proposed a time-varying velocity prediction model by utilizing fully convolutional networks (FCNs) for fast dif-feomorphic image registration. In contrast to prevailing unsupervised learning-based registration methodologies, the proposed approach incorporated the Laplacian operator of LDDMM into the framework of FCNs for parameter learning. The evaluation of the proposed method involved the utilization of two distinct T1-weighted magnetic resonance imaging (T1w MRI) datasets. Comparative analyses were conducted against two conventional diffeomorphic image registration algorithms and two state-of-the-art FCN-based methods. Experimental results show that the proposed method performed significantly better regarding registration accuracy while also obtaining desirable diffeomorphic properties.