Bilateral Asymmetric Analysis In Mammograms

With the progress and development of society,people’s living standards has been improved significantly,and the lifestyle has been changed greatly.At the same time,since the incidence of various diseases are increasing,the drawbacks brought by the social progress should not easily be overlooked.In the female group,breast cancer has already been the primary malignant disease that threatens women’s health.The computer-aided detection system based on mammography usually includes state of the art algorithms in the field of machine learning and computer vision.CAD aims to design a method that automatically detects lesions such as masses in the mammograms.Results of the detection assist the diagnosis of patient from doctors,therefore detection accuracy rate of breast cancer and the survival rate of patients are improved greatly.The human being has an important physiological characteristic that the breast tissue grows symmetrically,but the diseased tissue such as masses grows and exists asymmetrically.In the clinical diagnosis process,the doctors also refer the left and right mammograms to make the diagnosis.Therefore,we propose an algorithm for the asymmetric analysis of bilateral mammograms in this paper,and make a deep study in asymmetric analysis and mass detection with the bilateral mammograms jointly.The main work is summarized as follows:Firstly,this paper presents a cooperative asymmetric analysis framework for bilateral mammograms,combining the information of bilateral mammograms,discriminating the region pairs in the mammograms for symmetric or asymmetric,and making a classification for the asymmetry regions.The framework mainly includes the matching process of bilateral mammograms,asymmetric analysis based on sparse auto-encoder networks,and mass classification based on linear Support Vector Machine.Secondly,in order to achieve the alignment of bilateral mammogram’s corresponding structures,a shape context matching method based on adaptive matching point selection combining the structural features of the breast was proposed.In this method,the contour points of breast was extracted firstly,then matching points was chose adaptively with fixed intervals on the basis of bilateral nipples.In the matching process,the matching cost of bilateral points was calculated in log polar coordinates.The transform relationship between bilateral mammograms was estimated according to the best matching transform relationship returned by optimal matching,which could improve the accuracy and efficiency of matching.Finally,due to the difference of breast masses,an asymmetric analysis method of bilateral mammography based on multi-scale sliding windows and sparse auto-encoder network was proposed.The multi-scale sliding windows could make up for the lack of a single scale sliding windows,so that the masses could be wholly included in the block of mammograms,then the deep features of the image block was extracted.At the same time,the sparse autoencoder was introduced in the asymmetric analysis process.The SAE is used to learn the mapping relationship between mass-normal and normal-normal pairs in mammograms.The trained network can describe symmetry or asymmetry for the arbitrary input pairs,providing a comprehensive analysis for doctors.Then,the linear Support Vector Machine is introduced to classify the asymmetric region pairs to be mass or not.The experimental result shows that the algorithm under the proposed asymmetric analysis framework performs well on asymmetric analysis of bilateral masses and also obtains higher detection sensitivity.