ABSTRACT
Validation of image registration is a challenging task. The vast majority of image registration algorithms rely on models of human tissue that often deviate substantially from the true biomechanics. Even when biomechanical models are employed, the algorithms must still rely on other approximations and models. The combination of these approximate models with the reliance of the registration algorithms on the images themselves (i.e., the accuracy of the registration algorithm is tied to the complexity, clarity, and uncertainties in the image) leads to a difficulty in validating them. One can simplify the relationship between the two images to be registered (e.g., image a physical phantom or generate a mathematical phantom) and therefore know the “correct” registration result, or one can estimate the “correct” registration and use realistic, clinical images. This complexity in validation differs greatly from other aspects of radiotherapy, such as validation of dose calculation algorithms. Given these circumstances, a multistep validation approach is likely the best method to ensure an accurate and robust registration algorithm.
