Breast Bensity Segmentation And Quantization Of Classification Based On Dual Views

As of 2018,breast cancer has become the number one cancer incidence among women,becoming the primary enemy of women’s health.Medical studies have found that breast density is an important criterion for the evaluation of breast abnormalities.Mammary gland molybdenum target X-ray images because of its effective breast disease screening become the preferred found early breast cancer pathological changes,most of today’s public data sets are more retained in mammary gland molybdenum target X-ray images,mammary gland molybdenum target X-ray images at the same time in the process of filming prone to artifacts,single view cause glands rich overlapping prone to lead to problems such as the density of breast screening appear error,Therefore,multi-view and high-precision classification of breast density is needed in clinic.Comprehensive studies at home and abroad show that there are three main difficulties in improving the accuracy of breast density classification.Difficulty 1: how to remove the invalid region pectoral muscle in the inner and outer oblique views of mammogram images,and the insufficient number of mammogram images in the existing public data set.Difficulty2: How to solve the problem of poor segmentation effect of breast density caused by artifacts in mammogram images of molybdenum target.Difficulty 3: How to solve the problem of line classification error in mammogram of mammogram caused by gland overlap in single view.Based on the above three difficulties in classifying breast density,this paper carried out the following work..(1)Mammographic image preprocessing and data enhancement In the mammogram images obtained from INbreast Open database,only a part of the real breast region was found,and the invalid background region improved the reading of the characteristic regions of mammogram images.Medical research has found that breast density is an important evaluation criterion for breast abnormalities.However,the existence of background artifact areas and invalid pecs muscle in mammography has seriously affected the segmentation and classification experiments of breast density.Therefore,this paper made preparation for mammography before input by means of image preprocessing.The invalid background area was removed by binarization and simply connected region analysis method and the breast region image was clipped.Finally,threshold method was used to remove the breast muscle in the inner and outer oblique view.For deep learning,a large number of data sets are required,but the INbreast open database adopted in this paper has the problem of insufficient data volume.This paper expanded the data of 410 images to 3200 images,after which the model can learn more information about breast characteristics.(2)Breast density segmentation based on dual attention mechanismIn this paper,a U-Net deep convolutional neural network model based on dual attention mechanism is proposed to improve the segmentation effect due to the existence of internal artifacts.Firstly,the preprocessed mammogram images were input into the collaborative attention module,and the weight of each part of the input mammogram images was calculated using the attention mechanism,and the information of the region of interest was obtained in a multi-directional and large range.Secondly,in the up-sampling feature fusion process,the superficial feature map and the deep feature map processed by the collaborative attention module are processed through the attention gate,and the weight is generated and assigned to each part to further enhance the local feature.Finally,the feature map after attention gate processing is combined with the feature map after up-sampling,and then through the collaborative attention module,the whole feature strengthening in the upsampling process is maintained until the end of segmentation.The average segmentation index of the improved segmentation model in breast density was as follows: Dice correlation coefficient(DSC)was 91.8%,and intersection ratio(IOU)index was 85.8%.The DSC and IOU of dense breast type with high cancer rate were 98.4% and 96.8%,respectively.It is helpful for doctors to accurately judge the types of breast density.(3)Classification and quantification of breast density based on dual viewsAt present,most of the publicly available data sets are based on mammogram images of breast molybdenum target.When the glands are abundant,the single view is prone to the problem of breast overlap,which leads to the error of breast density screening.Therefore,multi-view and high-precision classification of breast density has become the clinical requirement.Firstly,the results of breast density segmentation in single view were obtained through the double attention module segmentation network in Chapter 4.Secondly,the cephalopods and the medial and lateral oblique positions were formed into a unilateral double view of the breast,and the percentage of breast density was obtained by means of the pixel-based breast density quantization method proposed in this paper.Finally,the results are classified and compared with the BI-RADS results.The experimental results showed that the double view classification greatly improved the accuracy of breast density classification,and the average accuracy of double view classification was increased by 2.93% compared with single view.Therefore,double view quantitative classification of breast density can not only solve the problems existing in single view,but also provide more accurate classification results for medical staff.