Research On Benign And Malignant Classification Of Mammography Images Based On Deep Learnin

With the increase of life pressure and environmental factors,the risk of women suffering from cancer increases year by year with the increase of age.breast cancer is one of the most serious diseases that affect women’s health,and early detection and reasonable treatment are essential for prolonging the survival period of patients.Due to the lack of medical resources and education,early prevention and treatment of breast cancer still faces great challenges.At present,the most effective screening method for breast cancer is mammography.Imaging doctors can find any abnormality in breast tissue by consulting mammography images,and initially determine whether there is any disease in the patient’s breast.Early screening relies on doctors’ clinical experience and professional knowledge,which is prone to missed diagnosis,misdiagnosis and other problems.Therefore,improving the efficiency and accuracy of doctors’ film reading and providing doctors with auxiliary advice through computer-aided diagnosis system are the key to improving the prevention and treatment of breast cancer.In view of the difficulty of mass classification and segmentation in breast cancer screening,this paper selects breast masses in breast mammography images as the research object.According to the characteristics of breast masses,based on the application of deep learning algorithm,combined with medical image processing,computer vision and other fields of technology,the main research work is as follows:(1)A breast mammography image segmentation model based on Double U-Net is proposed to address the issues of limited imaging technology and equipment,high medical image noise,blurry boundaries that are difficult to determine,and complex texture of breast tissue.Double UNet integrates the network architecture of U-Net with the advantages of the SE module(Squeeze and Excitation)and ASPP module(Atrus Spatial Pyramid Pooling).By encoding and decoding image features,it achieves deep and shallow semantic feature fusion of the network.The SE module performs weighting operations on the original feature mapping at the channel level,learns the weights between different channels,and improves the model’s performance and generalization ability,The ASPP module can simultaneously process multiple sizes of features in the same convolutional layer,thereby improving the model’s recognition and segmentation ability for targets of different sizes.In addition,Dice loss function is introduced to further help the whole segmentation network achieve better segmentation results.The breast mammography image segmentation model based on Double U-Net achieved an overall accuracy of 99.27% in predicting breast mass segmentation results on the CBIS-DDSM sampled breast mass dataset,while performing better in other accuracy indicators.(2)A breast mammography image benign and malignant classification method based on an improved deep residual network is proposed to address the problems of manual feature extraction,high professional knowledge,and long-time consumption in the process of breast mammography image classification,as well as the various fuzzy tissue boundaries and low contrast of the image itself.Combining the feature extraction ability of the Deep Residual Network(Res Net),the algorithm uses data enhancement and transfer learning methods to solve the problem of over fitting the model due to the scarcity of medical data.At the same time,it combines Inception network and residual network,uses Inception network to replace the convolution layer and pooling layer in the residual network,effectively reduces the number of parameters and speeds up the calculation process,and introduces Se LU activation function to further optimize the network performance.Experiments have shown that the improved deep residual network on the DDSM dataset can achieve higher classification accuracy(94.24%)and better generalization performance.