Learning-based deformable registration using weighted mutual information

Deformable registration of multi-modality medical image remains a challenging research topic. The incorporation of prior information on the expected joint distribution has shown to noticeably improve registration accuracy and robustness. However, direct application of the learned joint histogram makes the algorithm sensitive to the difference between the training data and the test image. This paper explores a more intrinsic intensity mapping relationship using normalized pointwise mutual information, and integrates the learned relationship into the conventional mutual information (MI) to formulate a weighted mutual information (WMI). We further derive a closed-form expression of the first variation of WMI for non-parametric de-formable registration in a variational framework. Experiment results show that the proposed WMI is more accurate and robust than MI, and is less sensitive to discrepancies between the training and test images, compared to the method in [1]. In addition, our prior can be learned from only a subset of the image, and can be object-specific.