Research On Medical Image Classification Method Based On Hypersphere Multi-class Support Vector Data Description

Breast cancer is a serious health hazard common malignancy cancer and it is alsoone of the main causes of death among the women all over the world. Since thepathogenesis of breast cancer has been not completely clear, early detection, earlydiagnosis and early treatment is vital to control the morbidity of breast cancer.Currently in the clinical diagnosis of breast cancer, diagnosis imaging is the mostsuitable diagnostic method for middle-aged women and breast X-ray photography isthe most common method for early diagnosis of breast cancer.This paper mainly study the theory of multi-classifier, support vector datadescription, kernel principal component analysis and MapReduce parallel computingframework from the perspective of data mining. In addition, both the key technologyand the main algorithm have also been presented in the field of medical image. TheHypersphere Multi-Class Support Vector Data Description (HSMC-SVDD) and theHypersphere Multi-Class Support Vector Data Description Based on MapReduceProgramming Model (MRHSMC-SVDD) are proposed and applied in classificationof mammography, respectively. The main work of this paper is as follows:1. Hypersphere Multi-Class Support Vector Data Description algorithmSome redundant data with high correlation will not only increase thecomputational complexity in constructing a classification model, but also may have abad impact on effect of the classifier. And the traditional multi-classifiers are basicallycombined by binary classifiers. But it is faced with many problems such as lowclassification accuracy and long training time when categories have been increased toa certain quantity. Aiming at these problems, the HSMC-SVDD algorithm has beenproposed in this thesis. The advantage of the algorithm is that the Kernel PrincipalComponent Analysis (KPCA) is integrated in the presented HSMC-SVDD algorithmto effectively reduce dimensions in the early stage of building the classification model.And then each category data trains only one Hypersphere One-Class SVDD(HSOC-SVDD) when building the mathematics model. Thus system overhead will beeffectively reduced. 2. Hypersphere Multi-Class Support Vector Data Description classifier used inmedical image miningIn this paper, the HSMC-SVDD algorithm has been applied in medical imagemining. Experimental results on the standard data set of studying breast X-ray images(MIAS) show that we have obtained good performance in either training speed orclassification accuracy.3. Hypersphere Multi-Class Support Vector Data Description algorithm basedMapReduce Programming Model and used in medical image distributedmining.When the training dataset is increased to a certain degree, the construction ofsupport vector data description training model will be a computationally intensiveprocess. With the increase of online diagnosis and clinical medical data, the efficiencyof the HSMC-SVDD classifier will be obviously reduced in multi-classification. Bystudying the parallel data mining, we have designed a novel parallel mining algorithm.The HSMC-SVDD classifier has been deeply designed based on MapReduceprogramming model. So Hypersphere Multi-Class Support Vector Data Descriptionalgorithm based MapReduce programming model (MRHSMC-SVDD) has beenconstructed in this paper and used in medical image mining. Experimental results onthe standard data set of studying breast X-ray images (MIAS) show that the speeduprate of the training model of the MRHSMC-SVDD classifier is closing to linearspeedup rate curve with the increase of the number of nodes, when the number ofnodes is less than a critical value. However, the speedup rate will reach a flat trend,when the number of nodes is more than the critical value. If huge data set were to beclassified, speedup advantages would be more apparent.