A Bayesian approach to level-set based image segmentation with a prior sensitivity analysis

Image segmentation is an important task in many image analysis applications where it is often an essential first stage before further analysis is possible. The level-set method is an implicit approach to the image segmentation problem which is particularly useful when the number of regions is unknown and where there is little information about likely region size and shape. It can also deal with a time sequence of images involving complex changes in geometry, such as merging and splitting of regions. In such situations high-level, object-based approaches are often infeasible or at least computationally demanding. This paper will present a detailed derivation of the level-set algorithm for image segmentation, based on the Chan-Vese model, which will then be re-defined in a probabilistic framework which allows a wider interpretation of model components. Further, it also gives an intuitive setting in which generalization of the model can be considered and where statistical model validation techniques can be applied. The performance of the proposed method is assessed using simulated data, and a sensitivity analysis is performed which investigates the stability of the solution to changes in user-specified model parameters. Finally, the approach will be applied to real data. The results demonstrate that the level-set method provides a useful approach and that the sensitivity analysis contributes an important assessment of robustness to subjective model choices.