Feature Extraction And Classification Of Cross View Mammography Images

Breast cancer is the second leading cause of female death,which seriously threatens the physical and mental health of women in China.Early screening,early diagnosis,and early treatment of breast cancer are the only effective ways to improve the 5-year survival rate of patients.The molybdenum target X-ray examination has three characteristics: simple and quick operation,local non-invasive for breast cancer patients,and good imaging of local breast tissue,so it is widely used for early screening of breast cancer.However,the imaging characteristics of breast cancer in different periods are quite different,and are susceptible to interference and influence by the density distribution of breast tissue,making manual reading and intelligent analysis prone to missed diagnosis and misdiagnosis.Although the artificial intelligence method improves the performance of X-ray image analysis to a certain extent,most of them are based on a single image or separately analyze the characteristics of X-ray images of different perspectives,ignoring the comprehensive recognition of cross-view Xray image content by doctors during clinical diagnosis.In order to fully utilize and mine the contextual information in the cross-view mammography image,and accurately identify whether the X-ray image contains the lesion area under various tissue density distribution.In this paper,according to the diagnosis process of the clinicians,the dual-view mammography are used to explore the potential correlation characteristics of the image content between multiple views,so as to better distinguish the accurate category of mammograms under different tissue density conditions.The main work results are summarized as follows:First of all,for the current classification network that uses down-sampled images as input,resulting in the loss of some detailed information in the image.A dual-view feature extraction network is proposed,which directly takes high-resolution dual-view mammography images as input.By introducing an attention mechanism to selectively extract important feature information that is more worthy of attention.Bidirectional LSTM network is used to capture the contextual correlation and semantic correlation between different views,effectively improving the classification accuracy of the network.On this basis,the convolution kernel used by the feature extraction module is single,which cannot effectively deal with the complex changes of tissue structure density and lesions in mammography,thus affecting the performance of the network.The dilated convolution with different expansion rates is designed to extract the multi-scale and multi-receptive field features of image context perception.And the bottom-up and horizontal connection mechanisms are introduced to multi-level cross-fusion fusion of shallow features and deep features,further improving the classification performance of the network.Finally,in order to make full use of the relationship between different convolutional layers in the feature extraction module,the fusion strategy of different hierarchical features is further improved,and a multi-level attention gate network is used.First,the channel attention module is used to weight the channels of the local features.At the same time,the compatibility score calculated by the local features and the global features is used to weight the space of the local features to target multi-level features of different importance,which further improves the classification network accuracy.The experimental results show that the cross-view mammography image classification method based on multi-scale deep learning framework proposed in this paper can effectively extract the rich feature information of high-resolution mammography images in different views.It has good adaptability,can accurately identify the abnormal information existing in the X-ray image,and has achieved a high classification accuracy,thus providing powerful technical support for assisting clinicians to make accurate diagnosis.