| Breast cancer is the most common malignancy and the second leading cause of death in women worldwide.Early diagnosis and detection of breast cancer can significantly improve the survival rate of patients,and mammography is a major screening tool for early breast cancer detection and diagnosis.The diagnosis of breast cancer relies on radiologists.However,this initial diagnosis is mainly based on the experience and professional skills of radiologists,so it is highly susceptible to the subjective influence of radiologists,which can easily lead to missed diagnosis and misdiagnosis.In addition,this diagnosis approach is also time-consuming,labor-intensive,and inefficient work.Therefore,to improve the efficiency and accuracy of diagnosis,the computer-aided diagnosis system came into being.It is helpful to improve the efficiency and accuracy of breast cancer diagnosis and detection by providing a reliable reference for radiologists.This thesis studies three virtual technologies in the computer-aided diagnosis system of breast cancer with deep learning technology on mammography,including:（1）Mammogram classification.Since radiologists are more concerned about abnormal mammograms,and the subsequent procedures in the computer-aided diagnosis system all process for abnormal mammograms,the mammogram classification technology,which is used to obtain abnormal mammograms,is a crucial procedure in mammogram analysis and processing.The size diversity of breast lesions brings an enormous challenge to the classification of mammograms.Therefore,this thesis proposes a Multi-Scale Attention-Guided Network（MSANet）consisting of a series of Multi-Scale Attention（MSA）bottlenecks.Each MSA bottleneck composes a Scale Aggregation（SA）unit that can aggregate multiple feature maps of different sizes and a Multi-Scale Attention Module（MSAM）that can re-weight the feature maps of different sizes.MSANet is validated on the most commonly used mammographic databases DDSM,which obtains the AUC value of 0.9363.（2）Breast mass detection.Since radiologists are more concerned about suspicious mass lesion regions in mammograms,breast mass detection technology which refers to obtaining suspicious breast mass lesions is an important procedure in mammogram analysis.Breast mass detection suffers from two enormous challenges: The number of mammograms is quite small and the breast masses only take up a small proportion of mammograms,which cause the class imbalance;some breast tissues are easily detected the mass,which the reason behind it is likely that the breast mass has similar visual patterns and the model can hardly capture sufficient contextual information.To solve the above-mentioned problems,the improved Faster R-CNN（i Faster R-CNN）is proposed.Compared to Faster-RCNN,i Faster R-CNN have two improvements: Res Net-50 is used as the backbone of i Faster R-CNN;The classification subnet and regression subnet are improved to capture sufficient contextual information based on the proposed ERFM（Expanded Receptive Field Module）.i Faster R-CNN is validated on the most commonly used mammographic databases INbreast,which obtains the recall value of 85.75%.（3）Breast mass segmentation.Since the shapes of benign and malignant masses are usually regular and irregular,respectively,it is an important procedure in mammograms analysis to use breast mass segmentation technology to obtain the edge of the mass.In this thesis,it is found in the experiment that the U-Net model has a relatively poor performance in segmenting small-sized masses,and the reason may be that the receptive field of U-Net is limited.Therefore,this thesis proposes a Selective Receptive Field Module（SRFM）,which consists of Multiple Receptive Field Moule（MRFM）and Multi-Scale Selection Module（MSSM）.The former can generate feature maps that have receptive fields of different sizes;The latter can select the suitable feature maps according to the input information.Based on SRFM,the Adaptive Receptive Field Network（ARF-Net）is proposed.ARF-Net is validated on the two commonly used mammographic databases（INbreast and CBIS-DDSM）,which achieves 86.1% and 85.75% dice scores,respectively. |
