A clustering guided deformable model for MRI segmentation

The proposed approach overcomes the shortcomings of previous deformable models in segmenting magnetic resonance images (MRI). In previous models, definition of the initial contour is left to the operator. In addition, previous models show poor convergence towards boundary concavities. The new method consists of the following steps. 1) An adaptive K-means clustering algorithm generates uniform regions with inaccurate boundaries for the tissues in the image. 2) Boundaries of the desired tissue are extracted from the regions generated in the Step 1. Certain points on the boundaries are considered as the vertices of the initial contour. 3) Since the results of Step 1 may not include detail boundary information in some regions, a correction algorithm using neighborhood information is applied. 4) Dynamic contour model of Lobregt and Viergever is applied, but in defining the external forces, a deflection in the radial direction is implemented in the manner defined by Prince and Xu. This uses the idea of solenoidal external forces to help track boundary concavities. By convergence of this step, through diminishing velocity and acceleration of all vertices, the procedure is completed. Experimental results show that the proposed method tracks the concavities quite well. In addition, the initial contours for tissues with closed boundaries are obtained automatically, thereby no initial contour needs to be defined by the operator.