Neuroimaging by positron emission tomography (PET) is a major research tool to investigate functional brain activity. Unfortunately, the analysis of these images is complicated by the variability of the data, low signal-to-noise ratios, and limited spatial resolution. Another major difficulty is that the shape and size of the brain varies from one individual to another. The purpose of this project is to develop image processing algorithms for the quantitative analysis of such data sets. At present we are emphasizing statistical approaches. In particular, we have designed and implemented a new multivariate approach for extracting the correlations between series of PET images and a set of external variables. We have also developed an original wavelet-based method for the statistical analysis of glucose utilization differences between subject groups. This technique provides a rigorous way of trading spatial resolution for an improved signal-to-noise ratio, and appears to be more powerful than standard pixel-based approaches. We are also designing a general multiresolution procedure for the efficient registration of PET images. Our geometrical transformation model allows for translation and general affine transformations (2-D or 3-D), including scaling and rotation.
