CT-Colonography cleansing using Bayesian probabilities

Currently, virtual colonoscopy examinations require extensive bowel preparation because residual materials can occlude lesions or can be misinterpreted as polyps. Our goal is to investigate a probabilistic method to segment contrast enhanced residual materials and remove them from the rendering. The region around a sample position is modeled to contain mixtures of air, tissue and tagged intraluminal remains. For each image sample a probability vector is calculated expressing the probability that the materials of interest are present. A probability space is defined using the probabilities for pure materials as base vectors. Mixture vectors are constructed at 45-degree angles between the pure material vectors. The probability vectors are compared to the base vectors and the mixture vectors to classify them into material mixtures. Consider the layer between air and tagged fluid. Image intensities are similar to tissue. The scale at which the Gaussian averaged probability is calculated is increased until convergence: two successive scales result in the same classification. The Bayesian classification method shows good results with relatively large objects. However, edges of small or thin objects are likely to be misclassified: a too large environment is needed for convergence.

@InProceedings { STVPV03c,
  author       = "Serlie, I.W.O. and Truyen, R. and Vries, G. de and Post, Frits H. and Vos, F.M.",
  title        = "CT-Colonography cleansing using Bayesian probabilities",
  booktitle    = "Proc. ASCI 2003, 9th Annual Conf. of the Advanced School for Computing and Imaging",
  year         = "2003",
  editor       = "S. Vassiliades and L.M.J. Florack and J.W.J. Heijnsdijk and A. van der Steen",
  note         = "266--270",
  url          = "http://graphics.tudelft.nl/Publications-new/2003/STVPV03c"
}