Many image registration algorithms are formulated as optimization
problems with a gradient descent based solver, One difficulty with
designing and implementing such methods is in the implementation of
the gradient computation. This process can be time-consuming and
error-prone. In addition some functions do not have gradients that can
be expressed in symbolic form. Automatic differentiation is useful for
computing gradients of complicated objective functions. It moves the
burden of computing gradients from the programmer to the computer. So
far, AD has not been exploited for use in image registration. This
paper describes a software library the authors have developed to
automate the process of computing gradients of registration objective
functions. This can alleviate the job of registration designers
somewhat and potentially make it easier to design better registration
algorithms.