Segmentation of X-ray CT images using stochastic templates

X-ray computed tomography (CT), a non-invasive imaging technique, is being used increasingly in sheep breeding. Currently, considerable human intervention is needed to segment images into different tissues. This is undesirable because of its subjectivity and tediousness. We propose the use of deformable templates to automate the segmentation. A stochastic model has been constructed using a training set of 99 manually-segmented images: Fourier coefficients were used to parameterise the template boundaries, and the coefficients were reduced in dimensionality using principal components. As a matching criterion between a template and an image, a weighted sum of squares of the difference between pixel values and their expected values was identified using the training images. Finally, the Nelder-Mead algorithm was used to optimise the matching criterion in order to fit a template to a specific image. The results have been validated on an independent set of 99 images, and boundaries were positioned to an average accuracy of 2.7 mm