Sensitivity to Starting Parameters and Statistical Modeling

For registration methods that must employ an iterative search algorithm to optimize some cost function, it is possible to evaluate how consistent registration results are as a function of the starting parameters used to initialize the search. This strategy gives an estimate of the accuracy with which the search optimizes the cost function. For registrations that involve least squares minimization, the accuracy with which the minimum has been identified can even be computed on the basis of a single optimization, subject to some fairly restrictive statistical modeling assumptions [7]. Although inaccuracies in identifying the true minimum of the cost function will contribute to registration inaccuracies, it is important to recognize that registration accuracy may be poor even when the minimum of the cost function has been identified extremely accurately. This is true because the cost function is often only a proxy for true registration, so that optimal registration does not necessarily coincide with the cost function minimum. Cost functions that are based on image intensities, rather than distances associated with image features, are especially likely to have minima that do not coincide exactly with correct registration. Nonetheless, these methods may sometimes be helpful by showing that poor identification of the cost function minimum places an upper bound on the accuracy that a method can achieve. However, in many instances, the minimization strategy is extremely effective, and this bound involves errors so small that nothing meaningful can be inferred about true registration accuracy.

Was this article helpful?

0 0

Post a comment