The most important aspects of fMRI data analysis, as used on the data in this thesis have been presented above. The pre-processing steps improve the ability of the statistical analysis to detect activations. The choice of an appropriate statistical test depends on the assumptions that can be reasonably made, and the questions that need to be addressed. Decisions about the statistical significance of the results must take into account the multiple comparisons problem, and the reduction in degrees of freedom resulting from smoothing. Finally the results must be presented in an appropriate form, to enable the comparison of results between subjects.
In the whole process, many informed decisions need to be made, and it is difficult to propose a single 'black box' approach to the analysis, where data is fed in and an activation image is automatically fed out. However in order to formalise the analysis procedure to some degree, a standard analysis protocol was written. This is included as Appendix D.
There are many more issues in the analysis that have not been mentioned here, and much work that can be done. For example the Fourier Analysis methods for detecting periodic activations[32] , principal component analysis[33], clustering techniques which detect clump together pixels having time courses which vary in the same way, and nonparametric tests[34]. There are other methods for solving the multiple comparison problem [35] and for carrying out inter subject comparisons. It is this last area which requires the most attention, since, as fMRI becomes more and more a tool for neurologists to detect abnormality in brain function, the reliability of inter subject comparisons will become more important. It has been well shown that fMRI can give colourful pictures of the brain working, and it is necessary to take the techniques for data analysis onward if fMRI is to fulfil its clinical and neurological research potential.