Multi-modal volume registration using joint intensity distributions

Abstract

Abstract. The registration of multimodal medical images is an important tool in surgical applications, since different scan modalities highlight complementary anatomical structures. We present a method of computing the best rigid registration of pairs of medical images of the same patient. The method uses prior information on the expected joint intensity distribution of the images when correctly aligned, given a priori registered training images. We discuss two methods of modeling the joint intensity distribution of the training data, mixture of Gaussians and Parzen windowing. The fitted Gaussians roughly correspond to various anatomical structures apparent in the images and provide a coarse anatomical segmentation of a registered image pair. Given a novel set of unregistered images, the algorithm computes the best registration by maximizing the log likelihood of the two images, given the transformation and the prior joint intensity model. Results aligning SPGR and dual-echo MR scans demonstrate that this algorithm is a fast registration method with a large region of convergence and sub-voxel registration accuracy.

Deformable Model-Based Image Registration

“…With 15 randomly generated rigid transformations, we applied 3 different functions: bi-linear, Cubic B-spline and Gaussian, as the interpolation kernels to estimate the motion parameters. These transformations are normally distributed around the values of (10 • , 10 pixel, 10 pixel), with standard deviations of (3 • , 3 pixel 3 pixel) for rotation and translation in x and y respectively. Table 1 3 Segmentation-Guided Deformable Image Registration Frameworks [14] In this section, we address the second and third challenges we pointed out in the section 1.2: how to integrate segmentation and registration into a unified procedure so that the prior information embedded in both processes can be better utilized.…”

A Framework for 3D Polysensometric Comparative Visualization

“…In our research, registration algorithms are divided into two groups: 2D registration and 3D registration. We use mutual information [4,5] or FFT [6] to align our datasets in registration framework, in which the common part of two datasets is normally 2D. Also, we use 3D registration to fuse two 3D estimation models, I A V and I B V , in visualization framework.…”

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

“…In [5] the log likelihood of the joint intensity distribution of the observed images was maximized with respect to the expected joint intensity distribution. In [6] it was shown empirically that using the Kullback-Leibler divergence (KLD) was superior to the log likelihood measurement.…”