Reconstruction of time-varying objects in computerized tomography using a model-based neural network

A new method is presented for computerized tomographic (CT) reconstruction of time varying objects from parallel projection data. Such a method can be used, for example, to correct for motion artifacts due to patient movement or breathing during a CT scan. The basic component of the method is a model-based neural network, known as the maximum likelihood adaptive neural system (MLANS). Under the MLANS formulation, the object function is first modeled as a mixture of localized basis components, where each component is associated with parameters describing local object structure and dynamics of motion. MLANS then provides an iterative framework for optimizing the model parameters based on a measure of similarity between the model and the data. A bonus of the method is the flexibility to process incomplete data, where projections are available only over limited viewing angles. Results of computer simulations are presented to illustrate the algorithm performance