Maximum likelihood contour estimation using beta-statistics in ultrasound images

In this paper, we address the problem of detecting the contour of objects in speckled ultrasound images. A prior hypothesis is made on the first order statistics of the image intensity that is related to the shape of the objects: the mean and variance of the images can he non-stationary, but they should be invariant to the normalized distance from a particular point in the image to the contour center. Furthermore, the first order statistics are modeled with a beta distribution. The choice for that distribution is also validated with some experiments and seems to be more realistic than the log-compressed Rayleigh distribution reported in the literature for ultrasound images. These hypotheses can be validated for particular images by means of the hypotheses tests that we also propose in the paper. The algorithm is initially tuned to a particular image class by estimating the model parameters from a prototypical image of that class. Once our model is set, the contour is estimated in test images using the maximum likelihood (ML) criterion. The ML estimator is adjusted so as to give some means of contour regularization.