The aims of this study were (1) to describe statistically the fluctuation of the goodness of automated CT-MRI registration method (2) to evaluate a numerical parameter, scaled to [0,1] interval (lambda), for characterizing the population level accuracy of any automated CT-MRI registration algorithm on voxel similarity basis. The population level distribution of cross-correlation values between the reference T1-weighted images and the automatically registered images were investigated in five patient groups (brain metastatis, cavernoma, cranial nerve schwannoma, meningioma, trigeminal neuralgia). The evaluated distributions appeared as the mixture of two Gaussians and a peak at the 1.0 value. The evaluated distributions appeared as the mixture of two Gaussians and a peak at the 1.0 value, therefore we classified the result of automated registration into three accuracy types (AT), AT1: cross-correlation equals to 1.0, AT2: when the automatically registered image slightly differs from the reference one, cross-correlation ∼1.0, and AT3: when the cross-correlation is about 0.4. Pauto was introduced as the ratio of well fitted automated registration relative to number of all the registrations, Cupper and Clower are the mean of AT2 and AT3 distributions. The λ=Pauto*Cupper*Clower product was used as the measure of the goodness of automated image registration procedure at population level. The evaluated lambda parameter will be used to control the impacts of software modifications and to optimize the functional parameters of the evaluated preprocessing steps.
