Analysis Of Image Features And Diagnosis Of Benign And Malignant Of Breast Tumors In FFDM

Full-field digital mammography(FFDM)is a screening tool for breast cancer detection by getting a richly contrasted view of the breast.However,radiologists remain have difficulty in the accurate interpretation of FFDM through the human eye,and the false-positive mammograms are still a common occurrence in the breast cancer screening programs.Computer-aided diagnosis technology(CADx)has great potential in predicting treatment outcomes concerned with high-throughput extractions of large amount of quantitative image features.The CADx can improve the specificity of discriminating malignant from benign lesions in FFDM.The traditional quantitative image features are handcrafted in advance,and significantly reflect expert knowledge.Deep learning method can automatically learn the features and expand the method of feature extraction.Inspired by these technologies,in this study,we make the effort to build a new classification model for FFDM,and study the traditional handcrafted features(HCFs)extraction method and deep features(DFs)extraction method to develop effective measures of breast tumor and reveal the relationship between the image features and malignancy of breast tumor.The main contributions of this dissertation include:(1)Preprocessing of FFDM images.The mean filter,median filter,and Gaussian filter are compared and analyzed of the effect of noise reduction in FFDM image,and Gaussian filter is chosen to improve the quality of the image.The different discretization on FFDM image would effects the classification accuracy,and the optimal image discretization level is analyzed for the subsequent image feature analysis.(2)Features extraction of breast tumors in FFDM image.According to the diagnostic behavior and information of radiologists,23 quantitative HCFs are extracted from the two view FFDM images to quantify the images characteristics of breast lesions.All HCFs in this study are divided into two categories:11 Gray Level Gap Length Matrix(GLGLM)texture features and 13 shape features.The GLGLM measures local texture structure of breast lesion.The shape features reflects the difference of the benign and malignant tumor on shape.(3)The HCFs-based classification method in FFDM.The effects of discretization degree,feature selection method and classifier on image classification are assessed and compared.Then,a classification model through weighting multi-classifier is built for breast lesion.The result on comparison and MIAS dataset show that the multi-classifier classification model can reduce the false positives and the recognition ability of benign and malignant tumors based on shape features and GLGLM features also can achieve the desired effect.(4)The HCFs and DFs-based classification method in FFDM.DFs can describe global and high-level information of breast tumor,and the features method that combining HCFs and DFs contain more comprehensive tumor information.The experimental results show that,compared with the HCFs features method,the proposed features method can further improve the classification performance.