PDF version of this document

Alternatives

The main alternatives to non-rigid registration are rigid and affine equivalents, but these are merely impractical. In most real-world applications such as registration of brain-slice images, there is a very slim chance of getting satisfactory alignment of structures while preserving some continuity unless non-rigid transformations are applied. One may argue that affine registration should suffice, but what if parts of the brain expand beyond proportion? It therefore appears as if, from a registration point-of-view, no obvious alternatives are yet known. The ones mentioned above give the best performance yet and comparison with the closely-related active appearance models suggests that flexible deformation is mandatory, especially for bio-medical data. Nonetheless, one could argue that there should be more than just a single alternative to be looked at and many different aspects call for attention as the earlier parts explain. Here is a short summary that may help guide future endeavours^B.4:

1. Speed-up: The methods operate very slowly for most globally-driven approaches. A solution to this is desirable because not only would it stimulate more experiments and experiment feedback, but it would also make these methods usable and marketable.
2. Data extension: The simple existing bump which is generated in MATLAB needs to be extended, possibly by conversion to a smoother bump as the one described in the research of Davies and Taylor.
3. Lambda coefficient: In practice, when constructing an appearance model for registration's sake, an additional weight is assigned to one of two related components. The first component is associated with the reparameterisation curve and the second corresponds to data values, i.e. intensities. This weighting term, denoted by Lambda (symbolically ) in the objective function, essentially weighs appearance against shape and its value is subjective and dependent upon the problem. Experiments can find (and have found before in Smith's work) alternative solutions or better assignments for lambda.
4. Automation: It would be desirable to create a (compilable) system that copes with the full cycle of analysis without outside intervention and without any pre-existent data annotation. This relates to the strands of artificial intelligence and autonomous systems.
5. Generalisation: Many ad-hoc algorithms are currently used for group-wise registration. An more impressive system would deal with arbitrary data without compromise to the quality of the results.