Study And Application Of Bayesian Theory On Medicine Image Processing

The Bayesian belief network is a powerful knowledgerepresentationand reasoning tool under conditions of uncertainty.A Bayesian beliefnetwork is a directed acyclic graph with aconditional probability distribution for each node. The graphstructure of such networks contains node srepresenting domainvariables, and arcs between nodes representing probabilisticdependencies. It encodes probabilistic relationships amongvariables of interest, so it can readily handles situations wheresome data entries are missing or noise exists. In the last decade,the Bayesian network has become a popular representation forencoding uncertain expert knowledge for expert systems. Recently,researchers have developed methods for learning Bayesian networksfrom data. The techniques that have been developed are still newand evolving. Some organizations and magazines (e.g ISBA) whichfocus on Bayesian research appear.On the basis of understanding and analyzing the currentresearch state, research focuses and development trend in thedomain of Bayesian network,this dissertation focuses on method ofthe Bayesian network structure is regulated based on the basicdependency relationship between variables and dependency analysisand many kinds of Bayesian Classifier and Application on MedicineImage Processing. In conclusion, the main achievements of thisdissertation include:1. This dissertation makes a survey about the research onBayesian network, including the background, the current researchstate, challenging problems and development trend, etc.2. The method of the Bayesian network structure to be regulatedbased on the basic dependency relationship between variables anddependency analysis is discussed. This methos is demonstrated byfour steps: creating the most power tree, adding the first kindof edges, deleting the second kind of edges, and directing theedges.3. This dissertation discussed several popular Bayesiannetwork classier models, in which Naive Bayes originates inpattern recognition and depends on the conditional independenceassumption. Although this assumption is rarely valid in real world,its predominant and robust performance receive much attention. Theexperimental study comparing the Naive Bayes classier to otherlearning algorithms (including decision tree and neural networkalgorithms) shows that the Naive Bayes classier is competitivewith these other learning algorithms in many cases and that in somecases it outperforms these methods.4. This dissertation discussed the medicine image processingusing Bayesian Theory. The difficulties come from the differenceof intensity of all sorts of objectives, noise of background andasymmetry of intensity. We reconsolidate the fragmentizedobjectives using Bayesian method to get unabridged objectives.Image segmentation aims to classfy objects. Therefore, weshould try to extract the complete objects without impurities ornoises. In other words, the enough not beyond information shouldbe saved. In segmentation, a series of methods were implementedto obtain the aim mentioned above. First, you should abstract theedge of images in order to stand out the contour of objects. Thenyou should make the new image to bi-value image, just having twocolors--black and white, which will lose infomation, but will savea lot time. You will do a eroding to leach noises of background.According to a lot of studies, we found that twice dilating hadbest effect because most success of connect and least error ofconnect. Search the connective regions. Combine two regions withmaximum posterior probability of the same object to be the resultsof segmentation.5. This dissertation discussed that the objectives of themedicine image are classed by Naive Bayes classier. We select theNaive Bayes classier in view of effect, efficiency and cost etc.The experimental study shows that the Naive Bayes classier is allright.To establish a system of identifying all kinds of objects, wemust first confirm which characteristics should be measured. Theseattributes called character and the parameter value of whichcomposed eigenvector. We got 33 characters about the shape ordistributing of intensity. There are relativities in thesecharacters which will impact the performance and veracity. Eachcharacter has its own traits. So we had done the characterselection. The seventeen charaters selected became the finalcharater of classier.In practice, we select the charaters by man-made andmachine-made method. The small standard deviation shows the steadyof the charater and the distances of charaters show therelativities. The criterion of the selection is that the distancecloses to zero and the standard deviation is small.Naive Bayes classier had been designed based these 17 charatersby the rule of minimum error rate. The type of object is the theroot and every charater is the leaf. The classier needs theConditional Probability Table(CPT). The Conditional ProbabilityTable can be acquired by account statistic of the sample set. Theexperimental study shows that the Naive Bayes classier is allright.