Diffeomorphic Image Registration with Neural Velocity Field
Abstract
The paper introduces a way to compute anatomically correct, smooth transformations between slices of a 3D image. Originally taken from the medical field (i.e. MRI scans), this method perhaps allows an alternative approach to aligning point cloud data.
Introduction
Image registration provides the correspondence and non-linear transformation between pairs of images. Diffeomorphic image registration (DIR) offers additional properties such as smooth deformation, topology preservation and transformation invertibility.
The authors propose to use recent advances in learning-based methods for DIR through deep neural networks for efficient computation of transformation.
They specifically mention neural fields for their potential in modelling dynamic scenes, the exact task at hand. They propose to realize DIR by optimizing an implicit neural representation of a continuous velocity field. Specifically, they take the following steps:
Pretrain a convolutional neural network to predict an initial deformation (short inference time + simplification of search space).
Diffeomorphic neural velocity field can optimize a residual deformation specifically for each pair of images.
