A new framework has been presented for general image segmentation using statistical shape model enhanced level sets. These are represented as a linear combination of non-Euclidean radial basis functions (RBF). A prior for the level set is represented as a probabilistic map, created using the training data.
Using a non-Euclidean distance for the RBFs allows the incorporation of image features with more accurate results. Additional benefits found are that the level set evolution can be represented as ordinary differential equations, and so no reinitialization is required. Furthermore, the results are topologically more flexible. The results of experimental applications demonstrate these advantages, and show the improvement of outcome over regular level set approaches.
Citations: [·]