Research On Detection And Diagnosis Of Breast Mass Based On Deep Learning

Breast cancer is the most commonly diagnosed among women in the world.And it also has a high mortality,posing a serious threat to women’s life and health.With the popularization of early breast cancer screening and the continuous progress of medical imaging technology,more and more breast cancer patients can be detected in early stage and treated in time.Among them,mammography is the most common medical imaging technology in breast cancer screening,which is characterized by low cost and clear imaging.In clinical work,physicians’ clinical experience and visual fatigue caused by long hours of work may lead to deviation of diagnostic results.At the same time,there is a high degree of similarity between breast mass and breast tissue,which will also affect the diagnosis of physicians.Computer-aided diagnosis has shown broad research prospects and application value in the field of medical image analysis,and can provide physicians with a "second opinion" for diagnosis.With the introduction of deep learning models,computer-aided diagnosis algorithms have been improved in practical applications.Therefore,in this thesis,a deep learning based breast mass detection and diagnosis method is proposed to help physicians quickly find mass lesions and provide the corresponding benign and malignant reference basis.Inspired by the combination of two-view mammograms in clinical diagnosis,this thesis proposed two-view breast mass detection method and two-view feature fusion mass diagnosis method respectively.Specifically,the contributions of this thesis can be summarized as follows:(1)Detection of breast mass based on two-viewMammography images are usually taken from both craniocaudal(CC)view and mediolateral oblique(MLO)view.Typically,the same mass will appear on both two-view mammograms.In order to provide physicians with more accurate reference results,this thesis proposes a method for breast mass detection based on two-view mammograms.First,in the first step of the method,an improved Retina Net detection model is proposed to detect suspicious masses from two single-view mammograms.Secondly,through the two-view matching algorithm,the suspicious mass regions can be matched to determine the true positive mass regions,and the non-mass false positive regions are eliminated.Finally,the proposed method was verified by a five-fold cross-validation experiment,and the recall and false positive per image were 0.951 and 0.42,respectively.(2)Classification of benign and malignant breast masses by combining two-view featuresTwo-view mammograms contains more features,and these features are correlated with each other.In this thesis,in order to make full use of the correlation characteristics of two-view features to improve the diagnostic performance,a classification model combining two-view features is proposed.Firstly,the model is based on the results of the two-view breast mass detection model,and the detected area of the breast mass is cropped.Then,two separate branches of convolutional neural network were used to extract the features of the mammogram patches from two views respectively,and the features of the two views were fused through the gated loop unit to achieve the final benign and malignant classification.Finally,the performance of the proposed model was also verified through five-fold cross-validation experiment,and the accuracy,recall and AUC value were 0.947,0.941 and 0.951 respectively.