Unsupervised Medical Image Registration via Dynamic Adaptive Total p-Variation Regularization

The sliding motion of organ edges and smooth motion within the organ co-occur because of the complexity of organ motion during respiration. Many registration methods assume a globally smooth deformation field, ignoring the problem of discontinuities in the field caused by organ sliding motion. In this paper, we propose a new regularization method called dynamic adaptive total p-variance. It was combined with the unsupervised deep learning registration framework VoxelMorph. The proposed approach dynamically estimates the position and magnitude of organ sliding motion from the deformation field during network training and then maps the magnitude to the p-values of Lp-norm (1<p<2). This approach adaptively assigns different exponential values of Lp-norm to each voxel, preserving the smoothness of the internal deformation field and the discontinuity of the deformation field at the sliding interface. The proposed method was tested on a 4DCT dataset of the lung. It outperformed other methods, including smooth regularization and locally adaptive total p-variation, in terms of sliding motion correction and preservation of discontinuities in the deformation field.