Mass Detection And Diagnosis Of Breast Mammogram Based On Improved YOLOv4 Model

Breast cancer is the most common type of malignant tumor among women in the world,and its incidence rate ranks among the forefront of female malignant tumors in the world and the mortality rate is very high.In clinical practice,mammography images are often used for early screening of breast cancer.To accurately detect and diagnose benign and malignant breast cancer,radiologists must screen a large number of mammograms every day.However,only relying on manual work to accurately diagnose the mass faces huge challenges.Therefore,it is very necessary to assist physicians in making a diagnosis through technologies such as computer vision and deep learning.Modern computer science technology can provide physicians with a second diagnostic opinion from a scientific perspective,so as to assist physicians in making accurate and accurate diagnosis more conveniently and efficiently diagnosis.In view of the current mainstream target detection algorithms in the detection of benign and malignant masses in mammography images,there are few applications,low detection accuracy and slow detection speed,and breast cancer screening benefits from multi-view visual analysis of conventional mammography images.However,due to various limitations of technological development,the existing target detection models for breast cancer do not fully utilize multi-view information fusion analysis.This thesis mainly takes the breast masses in mammography images as the research object.Deep learning algorithms,aiming at the above problems,carry out research in the following two aspects:1.Mammography image lump detection model based on improved YOLOv4:In order to improve the accuracy and detection speed of target detection in mammography images,a tumor detection model in mammography images based on improved YOLOv4 was proposed.This method can efficiently perform both detection and classification of lumps in one framework.First,the detection model introduces a split-aggregated dual-channel(JAnet)residual structure to improve the backbone network of the model;Second,use depthwise convolution and point convolution to replace the standard convolution in the original YOLOv4 model;finally,the post-processing stage proposes The larger value is averaged.The experimental results show that using the DDSM(Digital Database for Screening Mammography)data set as the training set to train the detection model,and using the INbreast data set as the independent test set,the proposed improved YOLOv4-based mammography image mass detection model has the Recall value,Compared with the original YOLOv4 algorithm,the m AP value,FPS and AUC value are improved by7.3%,6.45%,5.9fps and 13.02% respectively.The overall effect of the model is better than that of the current mainstream target detection models,showing good robustness and effectiveness,and can play a role in computer-aided diagnosis for physicians in the clinical diagnosis of breast cancer.2.Based on the improved YOLOv4 matching classification joint network model:Breast cancer screening benefits from multi-view visual analysis of conventional mammography images.In clinical practice,the robustness of the target detection model can be enhanced by integrating multi-view information,and the detection accuracy can be improved.According to the characteristics of mammography image data acquisition,the potential information between the dual views is fully utilized to improve the accuracy,a joint network model based on improved YOLOv4 matching classification is proposed.First,after introducing the Mish activation function,the matching degree of the mammary cephalopod(CC)and mammary lateral oblique(MLO)view pairs of the mammography image is detected by the Siamese network,and the similarity between the masses is calculated to obtain the double The matching degree of the view-image pair,then extract the image area that meets the threshold range of the matching degree to pre-position the tumor roughly,and then input all the mammography images that may have the tumor into the improved YOLOv4-based mammography proposed in this thesis In the image mass detection model,the mass is precisely located and classified.The experimental results show that the detection accuracy based on the improved YOLOv4 matching classification joint network model is better than the latest similar target detection models by making full use of multi-view information.On the INbreast dataset,the Recall value of this model for target detection of mammography images reaches 93.22%,the m AP value reaches 92.88%,and the AUC value reaches 97.93%,which is better than the traditional single-class target detection method.Compared with the original YOLOv4 model,the Recall value is increased by 9.1%,the m AP value is increased by 9.05%,and the AUC value is increased by 13.72%.The joint network model based on improved YOLOv4 matching classification proposed in this thesis effectively reduces the missed detection rate and false detection rate in mammography image mass detection,which proves the complementarity between the two tasks of classification and matching,and shows that the dual-view matching pair The role of performance improvement in the mass object detection task.The method improves the localization accuracy and detection accuracy of the target detection model by providing accurate dual-view mass correspondence,and further guides clinicians to provide effective auxiliary diagnostic opinions when diagnosing breast cancer.