Research On Segmentation Methods For Brain Megnetic Resonance Image Based On Multi-Atlas

The human brain is the most complex information processing system we have been studying at present.Scientific research about the human brain has recently been considered one of the most exciting fields.Brain disease has the characteristics of high incidence,high disability rate,high recurrence rate and high mortality rate,and presents a trend of youth.Its harm to human beings is equal to the sum of common diseases,such as cancer,cardiovascular and diabetes.Brain disease has become the biggest cost in the global social medical burden.With the development of science and technology,medical imaging equipment is more and more applied to clinical diagnosis.Medical images can provide more and richer diagnostic information,but the huge information needs to be analyzed accurately in real time.Therefore,medical image segmentation becomes very important and necessary,especially the segmentation of important brain structures.This paper includes the following aspects:1)The accuracy of multi-atlas segmentaiton depends on the registration.At present,there is no registration method can achieve perfect correspondence,each registration method has its own advantages and disadvantages.In order to integrate the advantages of different registration methods,extracting more useful image information,we propose the concept of multi-method and multi-parameter registration.That is to use different registration methods to register the atlas to the pending image with different parameters,get the registrated label images.Each registration method can obtain different registration results with different parameters.That can increase the number of training samples,thus more prior information can be introduced.2)Brain structures,such as the hippocampus and amygdala have blurre boundary and low contrast in MRI,traditional methods can not achieve accurate segmentation.We combine the advantages of multi-atlas segmentation and Active Contour Model,propose a new multi-atlas active contour model.It can not only introduce the prior information of the multi-atlas,but also get the smooth and continuous boundary curve.The model contains three parts:the multi-atlas term,the data term and the smooth term.The multi-atlas term fuse the information of different atlas by using local similarity of the registered atlas as weight.The data term correct the errors in registration by using the local gray information of the brain MR images.In the process of segmentation,the smoothness of the segmented curve is guaranteed by using the smoothing term.3)Traditional image segmentation methods are often based on single pixels,this kind of feature can only reflect the local information of the image.For the brain structure with blurred boundary and low contrast,it is difficult to achieve the ideal segmentation.There are different brightness between different atlas,when we segment MRI by using multi-atlas directly,this leads to inaccurate segmentation.So we propose a new concept:structural information descriptor(SID),which can reflect the relationship between the local and the global.It can provide more structure and direction information with stronger identification and significance,thus the accuracy of the segmentation can be improved.4)We introduces the classification information of the image.We use the cluster center of the structural information descriptor(SID)as the dictionary element,the local linear encoding of the important structure in the brain MR image was used to obtain its local linear expression.Combining the labels in atlas,obtain the last segmentation.