 |
 |
Hi again, last one now. From me at least. > >> it is true that the registration algorithm have an impact on the >> results. If not people wouldn't bother keep coming up with new and >> better algorithms. > > It definitely affects the results but they are not necessarily > "better" because "better" is extremely difficult to define. If > there is no "correct" transformation -- even theoretically -- then > it becomes critical to define what our criteria are and I would > suggest, there should be some broad agreement in the field on these > criteria otherwise one can always invent a new algorithm and claim > it is better on some measure. > I think we can agree that when registering two images using a non- linear registration they will look more similar than if we had used a linear registration? But the problem there is that we are registering a proxy for anatomy, namely the intensities in some MR image. The question then is if by getting this proxy better aligned do we also achieve a better alignment of anatomy (barring the somewhat ambiguous definition of "alignment of anatomy")? One way that people have been trying to work that out is by manually and painstakingly label regions in MR images, the labelling typically being based on "expert based gross anatomical features" (i.e. typically no cytoarchitectonal data for example). One can then register the images (e.g. pairwise) using the proxy (i.e. the MR images without any labelling info) and then see what method produces the better agreement of labels. When doing that one typically find that the agreement is considerably better for non-linear registration. I myself have been using the NIREP data set for this purpose and found precisely that. I agree that this is still somewhat circular in that the information that is available for the person performing the manual labelling is the same as is driving the registration (i.e. the MR intensities). But I would still argue that a trained radiologist/anatomist will have lots and lots of prior knowledge that allows him/her to use "holistic" information that goes well beyond the actual information contained in those intensities (which is really limited to tissue type). > So my point was simply that one criteria for "better" is > reproducible across labs, scanners, software packages, etc. in part > because more and more groups are using meta-analyses or just big > analyses of data archived in large storage systems when in fact, the > preprocessing steps may be artificially introducing unwanted noise. I think this is where we have different opinions. There are two sorts of reproducibility here that I think it is useful to distinguish between. We use different packages (e.g. SPM vs FSL) 1. Given we have the same set of data, do we get the same results? 2. Given we use the same paradigm, but I scanned my students and you scanned yours, do we get the same result? For 1 I agree that affine is more likely to give us similar results, simply because there are less aspects in which the algorithms can differ. For 2 I would claim that non-linear registration is more likely to give more similar results since we get a better agreement between subjects and hence we will get a more well defined "mean location" (cf. if the standard deviation is smaller, so will the standard error be). I would also claim that the second question is the more relevant in most contexts. Jesper