Student: Sam Ellis
1st supervisor: Andrew Reader, King’s College London
2nd supervisor: Paul Aljabar, Imperial College London
The most common type of PET scan involves a single static or dynamic acquisition of data for a single subject, coming from two modalities (eg PET and MRI). However, there are cases where it can be useful to consider multiple PET data sets, including for the following specific tasks:
i) Estimation of either averages or differences of kinetic and/or functional parameters between two or more PET scans of the same subject (to assess the same subject under different conditions, for example before, during and after therapy, or to more optimally use data acquired at close time points where physiology is assumed to be the same)
ii) Construction of PET/MR functional and anatomical atlases (to assist in image analysis, automated region identification, advanced simulation studies and for defining reference images and averages).
At present, both tasks are handled by post-reconstruction methodologies: the data are reconstructed in a conventional manner (e.g. using iterative reconstruction such as expectation maximization (EM)) and after reconstruction of individual data sets, the analysis is conducted to estimate the relevant parameters of interest. These parameters of interest are, for example, the averages or differences in kinetic parameters across scans for longitudinal studies of a single subject, or the mean shape and intensity of a functional and anatomical PET/MR image derived from multi-subject data (an atlas).
The advance in this project is to estimate the parameters of interest directly from multiple raw Poisson-distributed PET datasets, exploiting the synergistic simultaneous MR anatomical information for regularization purposes. In this way more data is used and more accurate noise modelling is included, and as a consequence improved parameter estimates are expected, with current research in direct kinetic parameter estimation showing that benefits range from 10% to 50% reduction in variance. This will mean obtaining images of the averages and differences in kinetic and/or functional parameters which have lower variance than single scan data, as well as lower variance than conventional averaging and differencing of multiple scans. Likewise, improved atlases (improved spatial resolution) are anticipated, with possible exploitation of the synergistic simultaneous MR data to assist in finding the diffeomorphic transforms needed to find joint anatomical and functional atlases.
Maximum likelihood (ML) or maximum a posteriori (MAP) objective functions will be used and the development of computationally challenging multi-dataset tomographic algorithms will be necessary, which operate on exceptionally large quantities of raw data (i.e. at least double the normal dataset size, as at least two datasets are involved in each case). Using a MAP objective function will allow an obvious way of regularization for PET from the anatomical MR data.
It is important to note that the advance will be in direct estimation of these parameters from the data, and any issues arising from differences in physiology (for finding averages) will remain the same as for a conventional approach. However, one goal is actually to estimate the difference in kinetic parameters (rather than the average), so no confound of signal averaging/fusion is encountered. Current work with direct non-linear kinetic parameter estimation methods in PET indicates that variance at the voxel level in image estimates can be notably reduced (from modest 10% reductions, even up to 50% reductions), and hence at least some benefit is anticipated in this present project. Lowering noise in turn means that more of the PET scanner resolution can be exploited as less need for image smoothing is needed.