Dynamic Magnetic Resonance Imaging is a non-invasive technique that provides an image sequence based on dynamic information for locating lesions and investigating their structures. In this thesis we develop new methodology for analysing dynamic Magnetic Resonance image sequences of the breast. This methodology comprises an image restoration step that reduces random distortions affecting the data and an image classification step that identifies normal, benign or malignant tumoral tissues. In the first part of this thesis we present a non-parametric and a parametric approach for image restoration and classification. Both methods are developed within the Bayesian framework. A prior distribution modelling both spatial homogeneity and temporal continuity between neighbouring image pixels is employed. Statistical inference is performed by means of a Metropolis-Hastings algorithm with a specially chosen proposal distribution that out-performs other algorithms of the same family. We also provide novel procedures for estimating the hyper-parameters of the prior models and the normalizing constant so making the Bayesian methodology automatic. In the second part of this thesis we present new methodology for image classification based on deformable templates of a prototype shape. Our approach uses higher level knowledge about the tumour structure than the spatio-temporal prior distribution of our Bayesian methodology. The prototype shape is deformed to identify the structure of the malignant tumoral tissue by minimizing a novel objective function over the parameters of a set of non-affine transformations. Since these transformations can destroy the connectivity of the shape, we develop a new filter that restores connectivity without smoothing the shape. The restoration and classification results obtained from a small sample of image sequences are very encouraging. In order to validate these results on a larger sample, in the last part of the thesis we present a user friendly software package that implements our methodology.

Document Type


Publication Date