Bayesian Approach Based on Geometrical Features for Validation and Tuning of Solution in Deformable Models

A local deformable-model-based segmentation can be very helpful to extract objects from an image, especially when no prototype about the object is available. However, this technique can drive to an erroneous segmentation in noisy images, in case of the active contour is captured by noise particles. If some geometrical information (a priori knowledge) of the object is available, then it can be used to validate the results and to obtain a better segmentation through a refining stage. A Bayesian approach is proposed to evaluate the solution of segmentation performed by a deformable model. The proposed framework is validated for vessel segmentation on mammograms. For that purpose a specific geometric restriction term for vessels on mammograms is formulated. Also a likelihood function of the contour and the image is developed. Our model avoids manual initialization of the contour using a local deformable model to obtain an initial contour approximation.