Biomedical Engineering Reference
In-Depth Information
algorithms from a medical imaging perspective. Gerlot-Chiron and Bizais have
presented a unified description of existing registration methods [9]. Maurer and
Fitzpatrick later adopted a similar scheme when they reviewed the registration
algorithms within the neurosurgery context [9]. Van den Elsen
et al
. reviewed
and classified medical image registration algorithms [10]. Their classification
criteria have been augmented and detailed by Maintz and Viergever recently
[11]. In addition to various survey articles and topic chapters (e.g. Fitzpatrick
et al
. [12]), a monograph on image registration has also been published [13]. For
further elaboration, the reader is advised to refer to the original surveys, topic
chapters, and monograph.
The increasingly complex schemes for classification reflect the sheer amount
of literature on image registration methodologies. It is impossible for us to de-
tail these algorithms here. However, we do want to point out a recent trend in
image registration research and practice, i.e., the voxel property-based registra-
tion methods have become increasingly popular. Compared to other registration
algorithms, the voxel property-based methods using the full image content offer
several advantages: they work on the image gray-value without any prior data
reduction; they can be automated and the results are objective; they require no
segmentation and involve little or no user interaction.
To this category, various paradigms have been reported, including cross-
correlation in spatial or (Fourier) transformed domain, minimization of variance
of intensity ratios, minimization of variance of gray values within segments,
histogram clustering and minimization of histogram dispersion, minimization
of the joint histogram entropy of different images, and maximization of mutual
information, among many others. Studholme
et al.
[14] compared five similarity-
based algorithms and Fitzpatrick
et al
. [15] compared 16 of these algorithms. The
reports from various independent groups confirm that the mutual information
maximization approach to image registration is one of the most robust and has
superior performance.
Mutual information image registration was independently proposed by Maes
et al
. [16] and Wells
et al
. [17]. However, their development is a natural con-
sequence of early effect on the analysis of voxel value joint histogram (see
[18]). Hill
et al
. [19] used third-order moments of the joint histogram as well
as other measures to characterize the clustering of the joint histogram at reg-
istration. Collignon
et al
. [20, 21] used joint entropy as a criterion for registra-
tion, but reported that it had a small capture range, i.e., only when the initial
