Ultrasound Image Segmentation using Stochastic Templates

Abstract

Point distribution models (PDMs) are incorporated into Bayesian image analysis, thus combining two approaches to the fitting of stochastic templates. Manually segmented images are used to identify both a PDM and a likelihood function, leading to a posterior distribution from which inferences can be drawn. The methodology 1s explored and illustrated using 144 ultrasound images of sheep. A pseudo-likelihood is found to give better results than a likelihood based on the distribution of pixel values in the training images. Estimates of sheep fat and muscle depths are shown to be comparable in accuracy with manual interpretation of images.