Research On Multi-task Classification Method Of Breast Mass Based On Mammography

The continuous deterioration of people’s living environment makes the death rate of malignant tumors continue to increase.Breast cancer is currently one of the most common tumor types in women and the second largest factor in the death rate of malignant tumors in the world.Detecting cancer at an early stage is crucial for successful treatment,and mammography is one of the main tools to aid in early detection.Breast masses are a primary indicator of the presence of breast cancer,the accurate classification of breast masses is of great significance,which can assist doctors in the diagnosis of breast cancer and meet the needs of clinical applications.Therefore,in the current medical field,research on classification methods of breast masses has gradually received more and more attention.In this thesis,based on the mammography image,combined with the current popular deep learning method,multiple classification tasks of breast masses are studied,mainly from two aspects: single-task classification method and multi-task classification method.On the single-task classification method,in order to determine the feature extraction network that needs to be used in subsequent research,this thesis first compares the singletask and single-view classification effects of different classic classification networks on classification tasks such as benign and malignant breast masses,shape,and BI-RADS evaluation level.Next,this thesis proposes a single-task multi-view classification model,by using the correlation between the image features of the CC and MLO views of the same classification task to perform feature fusion of the image features of the two views,to improve the classification effect of breast masses.On the multi-task classification method,this thesis continues the idea of feature fusion,and also utilizes the correlation between different classification tasks of breast masses,and proposes a multi-task and multi-view classification model.This model fuses image features from different perspectives from different tasks through the message passing mechanism of the graph neural network to further improve the classification effect of breast masses.Based on the largest breast mammography data set CBIS-DDSM,this thesis experiments to verify the influence of different modules in the breast mass classification model on the classification results and compared the accuracy of the single-task multi-view classification model and the multi-task classification model on multiple classification tasks of breast masses.In addition,this thesis also compares our method with the current advanced methods,which proves the effectiveness and superiority of our method,and provides a useful reference for the research and practice of breast mass classification tasks.