HIERARCHICAL PROBABILISTIC IMAGE SEGMENTATION

A method is described which creates a hierarchy of segmentations based on the Bayesian combination of information from different sources. Using receiver operating characteristic curves we show how incremental improvements can be made by adding further information sources. In this application, the segmentation hierarchy is presented to a user to obtain the final segmentation. The hypothesis generation strategy is based on the Canny edge detector and a simple but efficient gap closing technique. The resulting edge strength and interpolation statistics are combined with region statistics and symmetry information using Bayes' rule. The work is applied to MR and CT scan images of the human brain; the statistics used are calibrated against manual segmentations provided by expert clinicians. The conclusions are twofold. From a theoretical point of view, it appears that the edge detection extracts most of the grey-level information from these particular images; thus the improvement obtained by adding region statistics is minimal. In practical use, the initial feedback from the interactive system is encouraging, and suggests that the hierarchical form of the output is well-suited to interactive use.