Phase mutual information for MRI and PET image registration: parameter selection of Log-Gabor filter

Phase mutual information as a similarity measure has been demonstrated to perform comparably to an intensity method even when there is no strong intensity relationship between images. The estimation of local phase is significantly affected by the choice of bandpass filter and its parameter selection. The effect of the parameter selection of Log-Gabor filter on local phase estimation and performance of registration has been studied. The Log-Gabor filter is mainly constructed using two parameters, the center frequency w(0) and bandwidth B. Forty-two different transformations for magnetic resonance imaging (MRI) and simulated positron emission tomography (PET) data were used in this study, with mean and standard deviation of the distance between the gold standard position and the computed position taken as an error measure. The results showed that there was a tendency for wider B to be required to generate clear phase images with increasing of w(0). However, wider B could not improve the performance of registration for a certain w0. Parameters (w(0), B) with the values of (0.1, 3) are recommended to obtain optimum performance. There were no large differences in the registration errors among noise levels of 5%, 10%, 15% and 20%, suggesting that phase-based registration is robust against noise. When using the Log-Gabor filter with parameters (w(0), B) of (0.1, 3), the registration errors were always smaller than those in intensity-based registration for both cases with intensity non-uniformity in MRI of 20% and 40%. These results suggest that phase-based registration can be used as an alternative to intensity-based registration even in intensity non-uniform MRI.