Gao Zhiyong, Gu Bin, Lin Jiarui
Institute of Biomedical Engineering, Huazhong University of Science and Technology, Wuhan 430074.
Sheng Wu Yi Xue Gong Cheng Xue Za Zhi. 2003 Sep;20(3):476-8, 503.
Image registration methods based on mutual information, including mutual information and normalized mutual information, have been accepted as the most accurate and efficient methods. But there are many fluctuations in the registration functions that hinder the optimization procedure and lead to registration failure in intra-modal registration. We found that besides the interpolation artifacts, the uncertainty of the changing of entropy with the changing of overlap also contributes to the fluctuations. The effect of interpolation artifacts can be eliminated, but it is difficult to eliminate the effect of uncertainty of entropy. Luckily, this effect is not significant in normalized mutual information. Normalized mutual information is more stable and robust than standard mutual information and its better performance and wider application can be expected.
基于互信息的图像配准方法，包括互信息和归一化互信息，已被公认为是最准确、最有效的方法。但在配准函数中存在许多波动，这阻碍了优化过程，并导致模态内配准失败。我们发现，除了插值伪影外，熵随重叠变化的不确定性也会导致波动。插值伪影的影响可以消除，但熵不确定性的影响很难消除。幸运的是，这种影响在归一化互信息中并不显著。归一化互信息比标准互信息更稳定、更鲁棒，可以预期其具有更好的性能和更广泛的应用。
