A Bayesian spatial model with informative priors, with application to cone-beam computed tomography
Presenter: Matt Moores
Cone-beam computed tomography (CT) is widely used as part of the radiotherapy treatment procedure for certain cancers. It produces a 3D image of the patient’s anatomy in the region where the radiation dose will be delivered. The objective is to detect changes in the size, shape and position of the tumour and nearby organs, verifying whether the dose is targeted at the correct location in the body. This is a difficult task because cone-beam CT is characterised by low contrast-to-noise ratio in comparison to other medical images. It would be helpful to identify the objects of interest by segmenting the image according to tissue type, but overlapping distributions of voxel intensity values are an obstacle to accurate segmentation. In order to overcome these problems, we incorporate other relevant sources of information into our statistical model.
The hidden Markov random field (or Potts model) is a spatial model of the image lattice. The tissue type assigned to each voxel is dependent upon the types that are assigned to its neighbouring voxels. This contextual bias reflects the expected homogeneity of tissue type between voxels in close proximity with each other. We include an informative prior for the mean and variance of each tissue type, based on the voxel intensities that were observed in a previous image of the same object. We also include spatial prior information derived from a segmentation of the previous image. This is incorporated into the model as an external field.
The accuracy of the model is tested by applying it to cone-beam CT scans of an imaging phantom. This object of known geometry and density mimics the properties of a variety of human tissue types, including bone, muscle and fat. We demonstrate that the inclusion of the external field prior results in a substantial improvement in segmentation accuracy.