REGISTRATION OF MR MR AND MR/SPECT BRAIN IMAGES BY FAST STOCHASTIC OPTIMIZATION OF ROBUST VOXEL SIMILARITY MEASURES/

This paper describes a robust, fully automated algorithm to register i ntrasubject 3D single and multimodal images of the human brain. The pr oposed technique accounts for the major limitations of the existing vo xel similarity-based methods: sensitivity of the registration to local minima of the similarity function and inability to cope with gross di ssimilarities in the two images to be registered. Local minima are avo ided by the implementation of a stochastic iterative optimization tech nique (fast simulated annealing). In addition, robust estimation is ap plied to reject outliers in case the images show significant differenc es (due to lesion evolution, incomplete acquisition, non-Gaussian nois e, etc.). In order to evaluate the performance of this technique, 2D a nd 3D MR and SPECT human brain images were artificially rotated, trans lated, and corrupted by noise. A test object was acquired under differ ent angles and positions for evaluating the accuracy of the registrati on. The approach has also been validated on real multiple sclerosis MR images of the same patient taken at different times. Furthermore, rob ust MR/SPECT image registration has permitted the representation of fu nctional features for patients with partially complex seizures. The fa st simulated annealing algorithm combined with robust estimation yield s registration errors that are less than 1 degrees in rotation and les s than 1 voxel in translation (image dimensions of 128(3)). It compare s favorably with other standard voxel similarity-based approaches. (C) 1998 Academic Press.