Learning Based Non-rigid Multi-modal Image Registration Using Kullback-Leibler Divergence

Abstract

Abstract. The need for non-rigid multi-modal registration is becoming increasingly common for many clinical applications. To date, however, existing proposed techniques remain as largely academic research effort with very few methods being validated for clinical product use. It has been suggested by Crum et al. [1] that the context-free nature of these methods is one of the main limitations and that moving towards context-specific methods by incorporating prior knowledge of the underlying registration problem is necessary to achieve registration results that are accurate and robust enough for clinical applications. In this paper, we propose a novel non-rigid multi-modal registration method using a variational formulation that incorporates a prior learned joint intensity distribution. The registration is achieved by simultaneously minimizing the Kullback-Leibler divergence between an observed and a learned joint intensity distribution and maximizing the mutual information between reference and alignment images. We have applied our proposed method on both synthetic and real images with encouraging results.

Learning-Based 2D/3D Rigid Registration Using Jensen-Shannon Divergence for Image-Guided Surgery

“…In addition, as a section of experiments in [6], the superiority of KLD over Mutual Information (MI) measure was demonstrated for 2D/3D registration. Learning-based method was further extended to 2D non-rigid image registration in [7] where the KLD w.r.t. a prior joint distribution was minimized together with the maximization of the MI measure.…”

Registration of Cardiac SPECT/CT Data Through Weighted Intensity Co-occurrence Priors

“…In the following, we are deriving a new registration method considering those aspects. The achieved accuracy of the proposed approach, that exploits prior information from cardiac SPECT/CT acquisitions, is compared to the accuracy of standard mutual information (MI) [10,11] and a general learning-based approach [7]. This work extends previous works with the focus of applicability.…”

“…Decreasing the influence of MI allows to smooth out its local optima while still keeping the feature of maximizing the mutual information that both images share. In previous work [4,6,7,8] the KL divergence is used to measure the dissimilarity of two distributions. In a discrete formulation, this can be written as:…”

“…Prior knowledge about the correct mapping within this area is important to ensure such a mapping in future registrations. A problem of the general approach in [7] is that information stored in the learned joint pdf is global and influenced by the size of the background in both volumes. Local alignments are driven by the global matching especially if the transformation model is also global.…”

“…The accuracy is achieved by incorporating weighted intensity cooccurrence priors about an accurate alignment of cardiac data. Using the joint probability distribution function (pdf) of previously registered image data as an intensity prior has been previously reported supporting for the accuracy of rigid as well as non-rigid multi-modal image registration by several research groups [4,5,6,7,8,9]. Hereby, an interesting energy minimization scheme of incorporating statistical priors was proposed as follows [7,8,9]:…”