原文链接;https://www.sciencedirect.com/science/article/pii/S1361841516301074
1. A scheme to classify image registration methods
(i) dimensionality (spatial or spatiotemporal 2D/2D, 2D/3D, 3D/3D),
(ii) nature of the registration basis (extrinsic, intrinsic, non-image based)(no longer valid),
(iii) nature of the transformation (rigid, affine, projective, curved),
(iv) domain of the transformation (global, local),
(v) degree of interaction (interactive, semi-automatic, automatic),
(vi) optimization procedure (parameters computed or searched for),
(vii) modalities involved (mono-modality, multi-modality, modality to model, patient to modality),
(viii) subjects involved (intra-subject, inter-subject, atlas),
(ix) objects involved (e.g., brain, heart, breast).
Add:
(i) pairwise(n=2 images) vs. groupwise(n>2 images) registration,
(ii)asymmetric vs. symmetric formulations.
subdivisions of some of the categories would be due. For example, the category of optimization procedures could be divided into continuous and discrete methods, and for the category of curved transformations one could consider distinguishing small-deformation (or: elastic) and large-deformation (or: fluidic, based on integration of velocity fields) methods.
2.Public databases
Most of these concern manually delineated segmentations of structures, which are intended for evaluation of image segmentation methods but may also be used for evaluation of registration approaches. For example:
(1)public data sets of segmented MR brain images as IBSR (http://www.nitrc.org/ projects/ibsr )
(2)LPBA40 ( http://www.loni.usc.edu/atlases/Atlas _ Detail.php?atlas _ id=12 ) have been used for this purpose in studies on evaluation of registration accuracy, see e.g. Klein et al. (2009) .
(3)We would, however, like to draw the readers’ attention to the study by Rohlfing (2012) , which shows that the approach of evaluating registration algorithms on the basis of image similarity and tissue overlap measures has severe shortcomings and hence should be used with caution.
Just a few data bases have been set up specifically for evaluation of registration methods, all concerning deformable thoracic image registration, and primarily aimed at registration of inspiration/expiration scans of the lungs. These annotated data sets are provided by:
(1)DirLab ( http://www. dir- lab.com ),
(2)POPI ( http://www.creatis.insa- lyon.fr/rio/popi- model ),
(3)EMPIRE10 ( http://empire10.isi.uu.nl ). EMPIRE10 was launched as an evaluation challenge in conjunction with MICCAI 2010.
(4)Grand Challenges repository ( http: //grand-challenge.org )
(5)Retrospective Registration Evaluation Project (RREP), set up by J. Michael Fitzpatrick ( West et al., 1997 ). It concerned an evaluation of algorithms for rigid registration of CT, MR and PET images of the human head, aimed at support of neurosurgical procedures. The gold standard was obtained by registration of markers screwed into patients’ heads (as part of the clinical protocol).
(6)he challenge was continued as the Retrospective Image Registration Evaluation (RIRE) project, and is hosted by Kitware since 2007 ( http://www.insight-journal.org/rire ).
3.Four developments
(1)Registration research has focused largely on nonlinear registration (or ‘curved’ registration, as it was called in the original article);
(2)Intensity-based (‘voxel-based’) registration has become the method of choice also in multi-modal applications. The increased use of mutual information as a similarity measure has played a prominent role in this process.
(3)Inter-subject registration has gotten a larger share in registration research and applications.
(4)Generic registration software packages：
ANTs ( http://stnava.github.io/ANTs ),
NiftyReg ( http://cmictig.cs.ucl.ac. uk/wiki/index.php/NiftyReg ),
elastix ( http://elastix.isi.uu.nl ),
registration modules of the ITK toolkit ( http://www.itk.org ).
(As an example of the popularity of these packages, we give some statistics on our own software toolbox elastix ( Klein et al., 2010 ). )