Multi-Modal Image Registration Based on Multi-Feature Mutual Information

摘要

Multi-modal medical image registration is an essential technology in the field of medical image computing. Mutual information (MI) is one of the most common multi-modal medical image registration method. However, traditional MI only calculates the correlation of the global intensity, but ignores the local and structural information. Therefore, it is not robust against noise and spatially varying intensity inhomogeneity. In this paper, we bring the chain rule of information into high-dimensional MI and propose a multi-modal registration method based on multi-feature mutual information (MfMI). In MfMI, the mutual information between a set of features of the target and moving images is maximized. The feature set includes image's original intensity and structural features calculated from local derivatives and Hessians. In the experiment, we verified the robustness and accuracy of our method with synthetic and real patients' data. The results show that MfMI is robust to noise and inhomogeneity and outperform common registration methods to achieve higher accuracy. The mean target registration error (mTRE) of our method were 2.33 compared to 2.88 and 2.50 of LMI and alpha-MI in MR-US registration, respectively.