Probabilistic Partial Volume Modelling of Biomedical Tomographic Image Data

PhD, Centre for Vision Speech and Signal Processing, University of Surrey

Supervisor: Dr Kevin Wells.

The partial volume effect is an imaging artefact associated with tomographic biomedical imaging data. Three-dimensional volumetric data points (voxels) enclose finite sized regions so that they may contain a mixture of signals which are then known as partial volume voxels. The limited spatial resolution of tomographic biomedical imaging data, due to the complex biomedical image acquisition processes, often results in large numbers of these partial volume voxels. Clinical applications of biomedical imaging data often require accurate estimates of tissues or metabolic activity, where many voxels in the data are partial volume voxels. Therefore accurate modelling of the partial volume effect can be very important for such quantitative applications.

I investigated probabilistic models that provided a generic mathematically consistent framework in which the partial volume effect was modelled. Novel developments included an improved model of an intensity and gradient magnitude feature space to model the PV effect; a novel analytically derived formulation of the ground truth (prior) description of the PV effect; a novel gradient controlled spatially regulated classifier that utilises Markov Chain Monte Carlo simulations; and a fully automatic brain isolation technique that identifies brain voxels in neurological MRI data.

Simulated partial volume data and data from anatomical (MRI) and functional (PET) biomedical imaging modalities were utilized to assess the classification performance of the partial volume models. The data sets included: an imaged PET/CT phantom provided by the Royal Marsden Hospital, UK; publicly available simulated MR brain data together with the associated ground truths from the Montreal Neurological Institute, McGill University, Canada; and 20 normal MR data sets from the Center for Morphometric Analysis at Massachusetts General Hospital, USA. The performance of the developed classifiers were found to be competitive and in some cases superior to existing published quantitative estimation techniques.