Multi-Features for Mutual Information Based Medical Image Registration

Recently mutual information (MI) has grown to be accepted as a popular similarity measure and widely being recognized in the field of medical image registration. However, a known disadvantage of MI-based image registration is the lack of concern on any spatial information and involves only the pixel intensities as an input feature. To avoid this we are looking at the possibility of building a simple method to incorporate supplementary spatial information into MI estimation. Along with this direction, we propose multi-features mutual information (MF-MI) measure to simultaneously use all information obtained from multiple features. One major contribution of MF-MI is the incorporation of multiple image features into MI algorithm, while allowing a flexible choice of the spatial information through weighting coefficients. The method is thoroughly investigated and its accuracy and robustness are evaluated on both simulated and experimental data. Through quantitative evaluations, the MF-MI measure is proven to handle a wide variety of registration difficulties. Using multiple spatial features, the proposed algorithm is less sensitive to the effect of noise and some inherent variations, giving more accurate registration. In cases where MI performs well, the registration results of MF-MI are similar and the global maxima do not alter significantly. As a conclusion, the efficiency of the registration function is improved by using texture parameters to classify information-rich areas, and weighting coefficient methods to incorporate spatial information compared with the existing MI method.