Medical image registration is an important component of many clinical analysis pipelines. Image registration has conventionally been approached using variational optimization; deep learning has recently gained interest due to speed of inference allowing near real-time alignment of medical images. In this work, we have developed a method that performs image registration with the speed of deep-learning-based registration techniques that uses a gradient-based optimization with a piecewise learning rate schedule, achieving state-of-the-art accuracy in two medical image registration datasets. The results show that discrete optimization, which can be slow and computationally expensive, is not necessary to overcome local optima.
pdf