An Automatic Method Of Brain Extraction Based On The Graph-cuts

Nowadays, with the development of technology and economy, the medicaltechnology has improved quite a lot. And then varied medical equipments appeared,which can be used to measure and display the morphological structure of organs in thehuman body. Doctors can make a diagnosis to analysis the health conditions of humanbody through the important medical information in medical image. Brain is the keyorgan in the human body and the registration for brain images from different imagingdevices is particularly significant. Brain extraction plays an important role in thepre-processing of registration. Consequently, developing a fast and robust method ofautomatic brain extraction is conforming to the practical requirements.In recent years, with the advantages of fast optimal velocity and admirableextendibility, the graph-cuts has been widely used in the fields of image segmentation.So we developed an automatic method of brain extraction based on the graph-cuts,which is mainly applied to the T1-weighted MRI. As a kind of optimization methodbased on the graph, graph-cuts has the global optimality. But the cerebral MR imageexists particularity, it have no big gray level differences and uneven distribution amongthe tissues. Moreover, the traditional graph-cuts method is usually sensitive to theoriginal parameter setting and easy to cause problems of boundary leakage and localconvergence. Therefore based on the traditional graph-cuts, this paper proposed animproved method and adopted a new table of edge weights. Thus the accuracy of brainextraction is greatly improved.In order to ensure the robust of brain extraction, an Active Contour NeighborhoodModel is used in this paper. It adopts the improved BET (Brain Extraction Tool)algorithm to get a rough brain boundary. Then, the initial CN was obtained byexpanding the rough contour. After which the renewed contour line and Active ContourNeighborhood is got by the improved graph-cuts step by step. Finally, the accuratecerebral contour is obtained. The improved BET is fast and has semi-globalconvergence, and the resulting initial ACN is very close to the true brain boundary.Besides, the active contour neighborhood model is iterated and updated constantly, so ithas certain self-adaptability.In addition, for3D MR image sequence, this paper uses a slice by slice method ofthe2D slice. Due to the continuity of the brain surface, the initial contour for the middle slice was estimated by the modified BET. But for the rest of the slices, the initialcontour was estimated by the basic change of its adjacent slices. The initial ACN of the2D slice which obtained by this initial method is more accurate, and it can reduce theerror of the subsequent brain extraction process. In this initial method, BET isperformed only once on the middle slice, so it can save computational time.In our experiments, all of the MRI data used in this paper is from the Internet BrainSegmentation Repository (IBSR), the data base for researchers to evaluate the brainextraction algorithm which is developed by the Center for Morphological Analysis atMassachusetts General Hospital. Using the38samples of public sequence and itscorresponding manual segmentation standard provided by the IBSR data, the results ofour method were compared with the standard segmentation and the related evaluationparameters of the brain tissue segmentation were computed. The results indicate that ourmethod is better than some existing methods. So it can demonstrate that the proposedmethod in this paper is feasible.