Research On Contrast-enhanced Spectral Mammography Classification Method Based On Attention Mechanism And Information Bottleneck

Breast cancer is one of the most common malignant tumors in the world,which poses a serious threat to women’s life and health.At present,early screening using medical imaging means such as ultrasound imaging,magnetic resonance imaging,mammography and contrast-enhanced spectral mammography can reduce the mortality of breast cancer risk groups.With the development of medical imaging technology and the wide application of deep learning technology,the application of intelligent auxiliary diagnosis technology based on deep learning in early screening of breast cancer can help clinicians find lesions in time and effectively improve the diagnostic accuracy.Contrast-enhanced spectral mammography is a new screening method based on traditional mammography by injecting contrast agent.Radiation from head and tail position and lateral and lateral oblique position can simultaneously generate low-energy image similar to traditional mammography and dual-energy subtraction image showing abnormal vascular proliferation in tumor tissue.At present,researchers have carried out a series of studies in the field of contrast-enhanced spectral mammography image classification.However,most of the existing work only classifies breast cancer according to the presence of the tumor,ignoring the importance of the edge and shape characteristics of the lesion area in the clinical diagnosis of benign and malignant breast cancer.In addition,compared with traditional methods,the images obtained by contrast-enhanced spectral mammography contain more images and more comprehensive lesion information.However,most of the existing classification methods only focus on single-energy images or single-orientation images,and do not make full use of the complementary information in all images corresponding to each breast.In addition,there are some feature information irrelevant to the classification task in the image.Therefore,based on the deep learning technology,this thesis uses the residual network as the basic framework to study the image classification field based on contrast-enhanced spectral mammography.The main research contents and innovations of this thesis are as follows:(1)In order to solve the problem of edge feature extraction and preservation of the lesion area,a method of contrast-enhanced spectral mammography image classification based on coordinate attention mechanism was proposed.The network performs two one-dimensional coding operations along the horizontal and vertical directions at the same time,and integrates the feature vectors with embedded specific direction information into the original feature map.While accurately locating the location of the tumor in one direction,the network can capture the context information between the lesion area and its adjacent breast tissue in the other direction,which makes the classification network pay more attention to the edge features of the lesion and improve the classification accuracy.(2)In order to make full use of the complementary information between multiple examination images,a new method for classification of multi-input contrast-enhanced spectral mammography images based on context features and information bottlenecks is proposed.After the feature map containing context features is extracted through the backbone network,the network calculates mutual information and obtains the potential representation of multiple feature maps relative to the classification label.We transform and deduce according to information bottleneck theory and variational reasoning.And we use a decoder and an encoder to optimize the network parameters,so as to make reasonable use of the complementary information in multiple images and reduce the false positive rate of diagnosis.To sum up,the proposed classification method of contrast-enhanced spectral mammography image based on coordinate attention mechanism and the classification method of multi-input contrast-enhanced spectral mammography image based on context features and information bottlenecks were verified,and Py Torch was used as the basic framework to carry out the experiment.We tested on the contrast-enhanced spectral mammography data set,and evaluated the performance of the classification method proposed in this thesis through qualitative analysis and quantitative evaluation.Ablation experiments were carried out on key parts to fully verify the feasibility of the classification method.