Classification Of Breast Density Based On Multi-classifier Competition

Breast cancer is one of the common killers of women’s health all around the world,and it has increasingly caught the attention of the international community.However,the main exact cause of breast cancer is not given by the experts in the medical field at present,only the risk and predisposing factors have been analyzed and discussed.Breast density as an indicator of breast abnormalities,has aroused great attention from the medical groups.The goal of this paper is to explore and study the classification of breast density method based on mammography X-ray image that is the first choice for screening of breast diseases,aimed at improving the accuracy of early screening of breast diseases.After exploring and summarizing the research status of domestic and foreign scholars in the classification of breast density,this paper proposes a novel method for classification of breast density based on multi-classifier competition,which mainly includes breast gland tissue extraction,gland feature extraction and classification of breast density,the work of this paper is briefly introduced as follows:a)Mammographic gland tissue segmentation.Definition of breast density : Regions of brightness associated with fibroglandular tissue are referred to as ‘mammographic density’.In this paper,we firstly segment the glandular tissue to explore the breast density classification method based on fibroglandular region instead of the whole image.In order to obtain accurate fibroglandular tissue regions,the label,chest muscles and background interference are removed;then,the morphological top-hat transformation and nonlinear strategy are used to highlight the bright gland region;finally,the Gaussian Mixture Model(GMM)is used to segment the glandular tissue.b)Glandular feature extraction.The mammographic gland tissue acquired by GMM inherits the particularity andcomplexity of mammograms.According to the gray information of images,the gray-level descriptors is calculated,including energy,entropy,homogeneity,maximum correlation coefficient etc.;to give the better representation of texture saliency for the glandular tissues,five kinds of texture descriptors such as roughness,contrast,directionality,linearity and coarseness are calculated;finally,the concept of density descriptor is proposed to characterize the proportion of normalized breast gland tissue relative to the breast region,which can be regarded as a kind of global feature.The three types of features above constitute a density-dependent feature vector for subsequent breast density classification task.c)Classification of breast density.A multi-classifier method based on BP neural network classifier,LVQ neural network classifier,RF classifier and ELM classifier is used.The experiments were separately carried out based on 209 normal images from MIAS database and 320 normal images from DDSM database,respectively achieving the accuracy of 91.67%and 83.75%,these results demonstrate the accuracy and feasibility of the proposed method.