An adaptive systems approach to image reconstruction with a positron emission tomography application

This dissertation deals with an application of artificial neural networks (ANN) to image reconstruction with an emphasis in positron emission tomography (PET). In principle, through artificial neural network systems, it is possible to enhance heuristic tomography methods by incorporating adaptionto the imaging process. An adaptive image reconstruc- tion system seeks to respond to the context of the operation of an instrument (e.g., a PET system) compensating for the rigid response of phenomenological methods.

In mathematical terms, the imaging process in PET is equivalent to the inversion of a discrete Radon transform. Many are the obstacles for heuristic (i.e., conventional) recon- struction methods. The recovery of an object from a limited set of profiles is an expensive and difficult computational problem. The Radon inversion algorithm is the central component of image reconstruction in tomography.

The study of an ANN implementation of the inverse Radon transform and its effects in image reconstruction are discussed herein. The computer experiments are performed with both measured and synthetic PET data. The geometry of the transformation is pixel-based, and the ANN of concern is the back-propagation ANN.