Parameter estimation with a Bayesian network in medical image segmentation

Parameter estimation is a hard problem for image processing and segmentation tasks. In the case of 3D medical images, the segmentation can be performed slice-by-slice, extracting the Region of Interest (RI) in each slice. This task can be accomplished by using edge enhancement methods followed by Active Contour Models, also called snakes. In this case, we must set for each slice several parameters such as thresholds, kernel size for spatial filters, snake parameters (tension, rigidity), etc., which can be a tedious and time consuming task. In this paper we present a new methodology to adjust the parameter values in the slice l + 1 from the extracted region in the slice l. For each slice, several sets of candidate parameter values - and consequently sets of estimated regions - are randomly generated, and a bayesian network is used to maximize the probability of the RI in the slice l + 1 given the RI in slice l. We have tested this methodology in a large medical image database.