In the absence of gold standards, simulations are sometimes used to estimate registration accuracy. A common strategy is to take real data and deform it using an appropriate spatial transformation model while simulating the addition of noise and other factors thought to be relevant in limiting registration accuracy. Simulations are most useful when addressing the question of how sensitive a registration method is to some particular aspect of the data. For example, simulations might be very helpful when trying to choose the optimum amount of smoothing that should be applied to images for intensity-based intramodality registration. The results of such simulations can serve a very important role in optimizing the performance of a registration method. However, in the context of validation, simulations have definite limitations that can make them overestimate or underestimate registration accuracy. Simulations are especially poor in the context of comparing different methods to one another. The limitations of simulations derive from the fact that they are based on models of reality and not on reality itself. These models may omit factors that limit registration accuracy in the real world, or they may overestimate the degree to which a limiting factor is actually present. The models used to create simulated data for registration necessarily include spatial transformation models, interpolation models, and models of noise. Registration methods typically also implement spatial transformation models, interpolation models, and noise models either explicitly or implicitly. If the two sets of models are congruent, estimated registration accuracy will probably be excellent, but this provides little assurance that actual performance will be as good. To the extent that the models are not congruent, any poor performance will be difficult to evaluate since it can be blamed on the discrepancy between models.

0 0

Post a comment