Alignment by maximization of mutual information
ICCV '95 Proceedings of the Fifth International Conference on Computer Vision
| Hi-index | 0.00 |
An evaluation of a new similarity criterion for interindividual image registration is presented. The proposed criterion combines intensity and gradient information from the images to achieve amore robust and accurate registration. It builds on a combination of the normalised mutual information (NMI) cost function and a gradient-weighting function, calculated from gradient magnitude and relative gradient angle values from the images. An investigation was made to determine the best settings for the number of bins in the NMI joint histograms, subsampling, and smoothing of the images prior to the registration. The new method was compared with the NMI and correlation-coefficient (CC) criterions for interindividual SPECT image registration. Two different validation tests were performed, based on the displacement of voxels inside the brain relative to their estimated true positions after registration. The results show that the registration quality was improved when compared with the NMI and CC measures. The actual improvements, in one of the tests, were in the order of 30-40% for the mean voxel displacement error measured within 20 different SPECT images. A conclusion from the studies is that the new similarity measure significantly improves the registration quality, compared with the NMI and CC similarity measures.