PDF version of this entire document
The next stage was to build on 1-D and apply a similar approach using 2-D image sets. The goal was to register in 2-D and also to build a model in 2-D. I personally helped in making such registration work, so here are some important steps taken along the way in order to build this type of models which are derived from brain data (see, for example, Figure apm_brain).
This section explores an extension of the image registration method to 2-D and later comes a derivation of models from the registered data []. It is important to emphasise that this one particular goal was achieved by a group of people, so it should not be counted as an achievement made solely by myself, the author. I primarily contributed towards making NRR work better in 2-D, but not the automation of model building in 2-D, which is derived directly from NRR. I studied various compositions of registration stages in the NRR algorithm where brain images are aligned gradually, using a coarse-to-fine approach that improves speed and produces better results. These registration stages start with rigid and affine transformations and later consider different optimisations, scales, iterations and so forth (when non-rigid registration takes place). I first dealt with pairwise NRR, to contrast with later experiments that were purely groupwise. I did not deal with various arbitrary pairs of images; rather, this was always pairwise to a single reference.
These experiments were all done in C++. I used empirical evidence to show which paths lead to better solutions and which ones do not. There will be further discussion of these points in the rest of this section.