Graph Cut Based Magnetic Resonance Image Segmentation

Magnetic resonance imaging (MRI) is an important supplementary means to heart disease diagnosis because of its unique advantage. However, there are hundreds of images taken from an MRI of a patient, which make Magnetic resonance image segmentation based on computer technology become an important research topic in medical image processing.But it is difficult to segment cardiac magnetic resonance image automatically as a result of its complexity. So scholars have put a lot of efforts in interactive image segmentation which needs user's control and guidance. Automated MR image segmentation is still a very difficult task.To address graph cut based cardiac magnetic resonance image segmentation, we present an improved algorithm, which is mainly composed of the following work:1) We choose appropriate pre-processing algorithm in order to acquire feature image. Cost function of our method is based on shape prior information and intensity of the input image, and an effective pre-processing algorithm which can lead to accurate edges is extremely beneficial for subsequent segmentation. The input image is preprocessed by anisotropic nonlinear diffusion filter, and then canny edge detector (with some minor changes made for our method) is used to acquire image edge information precisely. The edge image is called feature image.2) After acquiring the feature image, we segment the input image using graph cut based image segmentation algorithm with shape prior, and optimize the cycle detection algorithm in order to reduce time complexity.3) We present a multi-resolution image segmentation method based on Minimum Ratio Cycle, choosing Gaussian pyramid and wavelet transformation to implement multi-resolution analysis to both get better segmentation results and reduce time computation cost. Comparisons between the two multi-resolution analysis methods are also presented.Compared with the other methods, the proposed method can segment cardiac magnetic resonance image automatically. The proposed algorithm is also applicable to shape matching and video tracking. The shape prior information does not need complex training, and ensures that the algorithm can get the correct result regardless of loss and rupture of edge. Besides, the method doesn't need to adjust massive parameters. Experimental results show the method is effective.