Interactive segmentation method with graph cut and SVMs

Medical image segmentation is a prerequisite for visualization and diagnosis. State-of-the-art techniques of image segmentation concentrate on interactive methods which are more robust than automatic techniques and more efficient than manual delineation. In this paper, we present an interactive segmentation method for medical images which relates to graph cut based on Support Vector Machines (SVMs). The proposed method is a hybrid method that combines three aspects. First, the user selects seed points to paint object and background using a "brush", and then the labeled pixels/voxels data including intensity value and gradient of the sampled points are used as training set for SVMs training process. Second, the trained SVMs model is employed to predict the probability of which classifications each unlabeled pixel/voxel belongs to. Third, unlike traditional Gaussian Mixture Model (GMM) definition for region properties in graph cut method, negative log-likelihood of the obtained probability of each pixel/voxel from SVMs model is used to define t-links in graph cut method and the classical max-flow/min-cut algorithm is applied to minimize the energy function. Finally, the proposed method is applied in 2D and 3D medical image segmentation. The experiment results demonstrate availability and effectiveness of the proposed method.