Discrete Unsupervised 3D Registration Methods for the Learn2Reg Challenge

The Learn2Reg challenge poses four very different tasks with varying difficulty for image registration algorithms. In this short paper, we describe our choices for two state-of-the-art discrete 3D registration methods that enable fast and accurate estimation of large deformations without expert supervision during training. Both approaches primarily focus on the use of contrast-invariant features with dense displacement evaluation and were ranked among the top three of all challenge contestants, yielding two first places and three second places for the four sub-tasks.