Statistical image analysis in cone-beam computed tomography

• Main Author: Doshi, Susan
• Other Authors: Jennison, Christopher
• Published: University of Bath 2014
• Subjects: 616.075722
• Online Access: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.619218

Cone-beam computed tomography (CBCT) is used to verify the patient’s position prior to commencing radiotherapy treatment. Soft tissues such as the prostate are hard to distinguish, and so gold markers may be implanted. These markers cause artefacts in the 3D reconstruction. In this thesis, we apply statistical image analysis techniques to CBCT data, with two purposes: we estimate the marker locations (with an assessment of uncertainty), and create reconstructions with fewer artefacts. In our first analysis, we define a Bayesian statistical model for the projection data, encouraging local smoothness in the prior. We use estimates of the true projection images (generated using Markov chain Monte Carlo, MCMC) in a conventional 3D reconstruction. The results are visually superior to those obtained using a frequency-domain smoothing filter. In our second analysis, we model the markers as they appear in the projection images. We restrict our model to regions of interest generated using morphological analysis. We combine the information from many projection images to generate an accurate estimate of the marker locations in 3D space. This produces accurate estimates of marker location, but no accurate measure of uncertainty. Our third analysis uses a template model for the markers in 3D space, with a separate model for the patient tissues. In phantom experiments, we obtain accurate estimates of the tissue properties and marker locations. For practical computational reasons, we can only analyse a small volume of the patient. Artefacts in the reconstruction used to determine the tissue properties outside the volume of interest prevent the successful estimation of both the tissue properties and marker locations in patients, but we accurately estimate the marker locations alone, with estimates of uncertainty. Additionally, we process the projection images, removing the markers. These processed images can be used to generate reconstructions with fewer artefacts.