Jvx

FIGURE 8 Simulated image that cannot be registered by maximizing mutual information. A is an original image, and B an image to be registered to A. Note that B looks similar to A except for the presence of shading in the horizontal translation direction. The graph next to B shows how mutual information changes with horizontal translation of image B with respect to image A. The correct alignment corresponds to zero translation. The same maximum value of mutual information is obtained for any translation by an integer number of pixels. C is a random permutation of image B. Note that the plot of mutual information against lateral translation when aligning image B with image A is the same for this image as for image B. This figure is taken from Roche eiai. [30].

FIGURE 8 Simulated image that cannot be registered by maximizing mutual information. A is an original image, and B an image to be registered to A. Note that B looks similar to A except for the presence of shading in the horizontal translation direction. The graph next to B shows how mutual information changes with horizontal translation of image B with respect to image A. The correct alignment corresponds to zero translation. The same maximum value of mutual information is obtained for any translation by an integer number of pixels. C is a random permutation of image B. Note that the plot of mutual information against lateral translation when aligning image B with image A is the same for this image as for image B. This figure is taken from Roche eiai. [30].

Although mutual information has been more successful than previous algorithms at registering 3D tomographic datasets such as MR, PET, CT, and SPECT, this success has been primarily in the head and is mainly limited to rigid-body or affine transformations. There have been encouraging results demonstrating the applicability of mutual information to both 2D-3D registration and nonrigid deformation algorithms, but these results tend to be quite preliminary. No one has yet shown that mutual information is the ideal similarity measure for intermodality registration. Indeed, the widespread use of rather ad fooc normalizations of mutual information suggest that it probably is not ideal. It is likely that other similarity measures will continue to be proposed that can be shown to succeed where mutual information fails, and the problem of identifying a theoretically ideal measure remains unsolved.

0 0

Post a comment