Segmentation Algorithm Based On Markov Random Field Theory Of Brain Magnetic Resonance Image

There are three general goals in this application: extract the brain volume; segment brain tissue into gray matter, white matter and cerebrospinal fluid; delineate specific brain structures. The objective in this paper is to segment brain tissue into gray matter, white matter and cerebrospinal fluid.Methods introduced here include K-means clustering algorithm, fuzzy C means and method based on Gaussian mixture model. These three algorithms perform well unless the image is contaminated by severe noise, because these algorithms do not use the spatial relationship of neighboring pixels in the clustering process, which will result the noise is classified to a certain class as a normal pixel. MRFs model spatial interactions between neighboring or nearby pixels. And in medical imaging, most pixels belong to the same class as their neighboring pixels. So MRFs can be introduced to improve the robust of fuzzy C means and method based on Gaussian mixture model in this paper.The algorithms are evaluated with segmentation accuracy and degree of equality quantitatively, and illustrated throughout in simulated and real MR images. Conclusions can be drawn :1. When the noise nearly exists, all the five algorithms presented in this paper can be very effective. The segmentation accuracy is above 97%.2. When the noise is severe such as 9%, the segmentation accuracy of K-means clustering algorithm, fuzzy C means and method based on Gaussian mixture model decrease 4% averagely. But after MRFs is incorporated with fuzzy C means and method based on Gaussian mixture model, the segmentation accuracy increase 2.5% and 3% respectively, compared with those without MRFs.