Accuracy Study

In the case of accuracy we have shown that MCFLIRT optimisation routines, cost functions and sinc interpolation consistently achieve high levels of accuracy. In particular, the RMS test measure shows that the error is typically around 0.1mm which is more than an order of magnitude less than the voxel size of 4mm, but necessary to ensure subsequent statistical analysis is valid. Furthermore, tests demonstrated that the average performance of the MCFLIRT scheme was superior to both SPM99 and AIR v3.08. Results of the tests using real data that contained activation, where the underlying ``ground truth'' was not known and the RMS measure could not be used, were inconclusive due to the presence of (an unknown amount of) physiologically induced intensity variations. Note that in all cases (for the robustness and accuracy studies) the data sets used for comparative testing were independent of those used to tune the empirical parameters of the methods used. Early tests using synthetic data have revealed that in cases where the motion is moderate (up to 2mm translation and 2$^{\circ}$ rotation), the sequential initialisation (see Figure 6) scheme yields an improvement in the accuracy of motion estimates compared to one where no sequential initialisation is performed. Conversely, in cases where the amplitude of motion parameters were known to be high, there was no inherent disadvantage in making the assumption of an underlying smooth motion trend across timepoints. We would be interested to see how robustly the schemes perform over time-series of varying length. If at all significant, we might expect to see some impact on the MCFLIRT mean image registration scheme where a longer time-series might provide a more general and robust template image. At present there is no guaranteed advantage in using the mean template in addition to the standard correction schedule but one would expect it to play a more beneficial role in correcting extended time-series exhibiting moderate to low motion artefacts. The methods and results shown here are all for whole brain data sets, although the algorithms have also been successfully adapted to work with data sets containing very few slices (or just single slices) by restricting the transformations to be two dimensional, as the optimisation and apodizing methods also apply for these two dimensional registrations. However, when general three dimensional registration is required for images containing few slices, other approaches need to be used, such as those employed by some of the other packages tested here (where no global search is involved). Practical registration packages usually require the setting of certain parameter values. The methods introduced here also contain several configurable parameters such as image resolutions for the various optimisation stages. These values have been selected empirically, over a wide range of data sets, to be as robust as possible for general brain images. However, when dealing with particular data sets these general settings may not be optimal and so most methods allow these values to be changed, via configurable options. When using other packages for comparitive studies we have tried to select the best general parameter values (by consulting the appropriate documentation) but recognise that further improvement upon these results could be possible by careful selection of parameters. The tuning of registration methods is, at present, an undesirable necessity in many situations, which prevents easy automation. By using more robust algorithms this tuning of parameter settings can hopefully be minimised or avoided entirely. In summary, the FLIRT and MCFLIRT packages are highly robust and accurate, as has been demonstrated by the quantitative experiments and by qualitative feedback. These methods have now been used to satisfactorily solve thousands of registration problems, some using extremely different imaging modalities. The binary and source code distributions for the MCFLIRT and FLIRT packages are available for downloading from www.fmrib.ox.ac.uk/fsl.