Research On Deep Transfer-based Breast Mass Detection Technology Of Molybdenum Target Images

Breast cancer is a common malignant tumor disease with a high incidence in women.Early detection and early treatment are effective methods.Molybdenum imaging is one of the effective methods for early diagnosis of breast cancer.According to research,physicians can use computer-aided detection(CAD)technology to find lumps in molybdenum target images at an early stage,helping physicians improve their work efficiency.However,due to the large differences in shape,size and density of different lumps,detecting and classifying lumps is still a difficult challenge.The traditional methods for the detection and classification of lumps are cumbersome and mostly rely on feature description and threshold analysis.These methods have poor generalization ability and robustness.With the rise of deep learning,deep learning algorithms have shown powerful functions in pattern recognition tasks(such as object detection and speech recognition),which have been widely used in the field of images.Compared with typical machine learning methods that require manual feature extraction,deep learning algorithms have more powerful feature learning capabilities.At present,deep learning target detection frameworks are complex and numerous.There is not a unified evaluation standard for the detection of molybdenum target masses,so it is difficult to compare the performance of the methods.In this paper,by combining the current rapid development of deep learning technology,transfer learning technology and target detection technology,the research and analysis of several key technologies for detecting and diagnosing breast masses based on molybdenum target images are as follows:1.In view of the complex and endless framework of deep learning target detection,there is no unified evaluation standard for the detection of molybdenum target masses,so it is difficult to compare the performance of the methods.This paper proposes a study of feature extractor models for breast mass classification detection based on deep migration,migrating different feature extractors to different methods,and exploring the application of target detection systems in the classification anddetection of breast masses in the medical field with a detailed and fair experimental environment and settings.The trade-off and comparison of different methods in speed and accuracy,and the use of the idea of deep migration to explore the correlation between different feature extractors and different frameworks,as well as the impact on the performance of the framework,which is intended to help physicians in the choice of frameworks and guide.2.For the R-FCN target detection structure,the image lost a lot of detailed information and spatial location information during operation.There are misdetections and missed detections on the molybdenum target images with higher breast background and tissue complexity.A method based on parameter optimization of R-FCN for breast mass detection and classification was proposed.By adding the activation function RELU-6 to the one-stage of the Res Net101 feature extractor of R-FCN,the initialization parameters weight truncated normal distribution is used to optimize the parameters.After the optimization is implemented,the model can better mine relevant features and fit training Data,accelerate convergence and improve accuracy.In the parameter initialization phase,a regularization method is used,and an initialization method that generates a random number with truncated normal distribution is used to set the standard deviation of the mean value to ensure the stable transmission of data in the network.