An adaptive deterministic annealing approach for medical image segmentation

We present a stochastic model based technique that uses the concept of deterministic annealing to obtain a generalized solution to the nonconvex optimization problem encountered by many image segmentation techniques. Deterministic annealing [DA] is an elegant and useful tool for clustering and classification. This novel optimization approach works with the efficiency of a deterministic procedure and has been successfully applied to a number of combinatorial optimization problems. We demonstrate effective segmentation of simulated MR brain images and provide a quality measure for accuracy of classification. A generalized deterministic annealing procedure, which works tender a structural constraint of mass or density, has been utilized for this purpose. This method produces a hierarchy of solutions giving segmentation results from a coarse to a fine level. Automatic edge detection can be performed using these solutions that are at different degrees of coarseness. The procedure has been made more efficient by utilizing a new similarity parameter from the concepts of neuro-fuzzy clustering.