Detection methods and evaluation studies for functional neuroimaging

Brain activation studies based on PET or fMRI seek to explore neuroscience questions by using statistical techniques to analyze the acquired images. An increasingly wide range of univariate and multivariate analysis techniques are used to generate statistical maps in which local mean-signal activations and/or long-range spatial interactions may be detected. In this thesis we review, compare and evaluate some of the existing functional neuroimaging analysis methods and then propose two new directions for neural activation estimation.; Some of the most commonly used brain activation detection methods were evaluated using an artificial brain phantom derived from real neuroimaging study and ROC curves. The results of the study suggest that multivariate techniques may be preferred over univariate methods. The findings of this study also inspired the design of a new method based on a Reversible Jump Markov Chain Monte Carlo (RJMCMC) technique. The key idea in this approach is to model the shape of an unknown activation map by using some predefined elementary shapes. When compared to the other methods tested, this new method showed a significant increase in detection performance.; The second new direction we explored is a method of Independent Component Analysis (ICA) for multiple data sets. We propose a model that enables the separation of independent components generated by the underlying processes of interest from other unwanted effects. We show synthetic data example to illustrate the advantages of the method and also perform an extensive evaluation on real fMRI data using resampling framework known as NPAIRS. This evaluation showed better performance of the new multi-set ICA when compared to standard ICA and Principle Component Analysis (PCA) algorithms.; We also discuss effects of spatial misalignments, often encountered in multi-subject data sets, on the results of PCA and ICA analysis. We identify the situations in which these effects are negligible and confirm these findings by both synthetic and real data experiments.