Classiifcation Of Masses In Mammography Based On Multi-Classifier Fusion With An Improved Multi-Agent Algorithm

Breast cancer is one of the most common malignant cancers among the women and has aserious negative effect on physical and mental health. In the methods for early diagnosing breastcancer, mammography is considered as a good way for the early detection of breast cancer.However, the ability of human eyes in reading mammography is limited, computer aided diagnosis(CAD) system could improves the reliability of early diagnosis and reduces the human factors ofmisdiagnosis. It is a new way to increase the diagnostic accuracy of the early diagnosis of breastcancer.Classification of breast masses plays an important role in mammographic CAD system. Manyclassification algorithms have been proposed for mass classification in mammographic CAD system.They are neural networks, decision trees, k-nearest neighbor, fuzzy clustering and Bayesian, etc.,while each method has its limitations as well as preferences. A method often achieves good resultswith one data set, while with another data set data it is unsatisfying. In recent years, a lot of researchhas been done on the classification with classifier fusion method, many experiments show thatmulti-classifier fusion algorithm can improve the classification accuracy. Multi-Agent (MA)algorithm is introduced as a novel multi-classifier fusion method recently. Previous studies haveconfirmed that MA algorithm increases classification accuracy, compared with other traditionalclassification algorithms of breast masses. But when agents try to consult each other, it does notconsider the effect of different performance of agents. While it takes benign or malignant decisionsit also does not consider that the iteration may not converge and refuses to decision-making. Thus,this thesis proposed a multi-classifier fusion scheme with an improved Multi-Agent (MA) algorithmto identify the benign and malignant masses. It considered emphatically that individual classifiershave very different performances during the classification. Furthermore, when the improved MAalgorithm counted the classification result, not the classification label but the confidence wasutilized.The main contributions and innovations are listed as follows:1) Improved MA multi-classifier fusion algorithm bases on conventional MA algorithm, withthe introduction of the weight matrix, it considers that each single classifier has differentperformance to identify the benign and malignant masses. The corresponding results show that theclassification accuracy of improved MA algorithm is better than conventional MA algorithm, andalso has good stability.2) In improved MA algorithm, when the number of iterations reaches a certain threshold, the weight-average algorithm instead to make decision. Experimental results show that improved MAalgorithm can make sure the result convergence.3) When the number of single classifiers which make up of improved MA algorithm changes,the performance of improved MA algorithm has a little impact.4) With the technology of the multi-views, improved MA algorithm has better performance onclassification of the benign or malignant masses.