Tuesday, November 19, 2002
It is generally accepted in PET that reconstruction algorithms based
on statistical principles can provide better performance for certain
tasks than those based on analytic inversion formulas of the X-ray
transform. However, due to the ill-conditioned nature of the
problem, it is necessary to introduce some kind of regularisation.
It is then important to be able to predict the properties of the
reconstructed images in terms of the measured object and the
regularisation parameters, such that these parameters can be
optimised for a particular task.
This talk concentrates on a regularisation method where the image is filtered after every iteration of the algorithm. Analytic formulas for the resolution and noise behaviour are derived. These formulas are first used to show that these properties are in general object dependent, and then to derive various approaches to obtain object-independent resolution. Finally, we obtain a connection between this inter-filtering regularisation method and penalised likelihood reconstructions.
Dr. Thielemans has a Ph.D. in Theoretical Physics, and in particular in String Theory. He now works as a researcher at Imaging Research Solutions Ltd, Hammersmith, UK. His particular interest is understanding and characterising the performance of image reconstruction algorithms based on statistical principles, as applied to Positron Emission Tomography.