Dr. David Hawkes is director of the Image Processing Group in the Division of Radiological Sciences at the United Medical and Dental Schools (UMDS) of Guy's and St. Thomas' Hospitals in London. Together with his colleagues, he has been occupied since the early nineties with the complex study of automatic medical image registration, using the method of voxel similarity measures. During the annual conference of the Advanced School for Computing and Imaging (ASCI'98), last June at the Vossemeren in Lommel, he was invited to share the preliminary outcomes of his research with the participants of this three day event.
Before the UMDS research team in 1992 entered into contact with Dr. Roger Woods, a neurologist from the University of California in Los Angeles, they believed it to be impossible to accurately register radiological images from different modalities through the use of similarity measures that were calculated from voxel intensity values. Instead, the British scientists tried to match equivalent features, such as points, lines, surfaces or volumes which they identified in the medical images.
Dr. Woods however minimized the variance of intensity ratios across the voxels in the registration of two PET images, originating from the same patient and succeeded in modifying the algorithm, in order to register PET brain images with magnetic resonance (MR) images, after the scalp and all the other extracranial material had been removed. This brought Dr. Hawkes and his team to the idea of applying the new technique to the registration of MR with computed tomography (CT) images. He was also interested to find out whether a change in voxel similarity measures could possibly enable the registration of MR and PET images without first having to remove the scalp.
The UMDS researchers decided to create so-called intensity feature spaces, comparable to those applied in segmentation classifiers. For this purpose, they selected images registered with a landmark based algorithm and looped through all the image voxels. As a next step, the intensity of a well defined voxel in one modality was plotted against that of the corresponding one in the second modality. The scientists were particularly intrigued by the change in appearance, observed in feature spaces in case of mis-registration. If they are normalized, they can be considered as a "joint probability distribution".
Maybe, the phenomenon of the feature space dispersion could be adapted for image registration, and the team tried to discover the ideal measure for the generation of a new algorithm. As a result, a formalism was introduced for the description of existing voxel similartiy measures with regard to the joint probability distribution. In addition, the scientists designed a robust multi-resolution optimization technique. Now, they were able to compare various parameters, such as the performance of Dr. Wood's variance minimization without scalp editing and several other voxel similarity measures, the correlation coefficient, and the moments as well as the entropy of the joint probalility distribution.
At that time, in 1993 and 1994, research teams at the KU Leuven in Belgium and at MIT Boston in the USA began experimenting with new voxel similarity measures for image registration, supported by the earlier UMDS findings. Independenly from each other, they came up with a more general measure of mis-registration, namely the maximization of the mutual information, equally referred to as "relative entropy", of the joint probability distribution. The British team immediately took up these results for use in their own optimization algorithm.
Whether mutual information is a better voxel similarity measure for MR-PET neuro registration than the modified version of Dr. Wood's algorithm without the scalp editing, remains to be seen, according to Dr. Hawkes and his team. Also, whether it forms the ideal measure for MR-CT neuro registration. In any case, this type of voxel similarity measure is able to accurately register both modality combinations. Please, consult the movie files and scientific papers, collected at the UMDS Web site for more detailed information.