| Breast cancer is one of the most common malignant tumors that harm women’s health worldwide.Mammography is the main imaging method of clinical detection of breast cancer,which plays an important role in reducing the mortality of breast cancer.With the continuous development of China’s economy and the improvement of women’s awareness of prevention and treatment of breast cancer,molybdenum target X-ray examination has been widely used in medical institutions at all levels in China.As the detection rate of mammography for breast cancer is about 0.3%-0.5%,it is a great waste of medical social resources for healthy women to participate in mammography examination every year;secondly,mammography examination has certain limitations and hazards,such as falsepositive results and radiation exposure.Therefore,in recent years,the social cost and effectiveness of mammography screening have attracted great attention from academic and medical circles.In order to reduce unnecessary molybdenum target census or reduce the frequency of census,it is very important to evaluate the population in the near future,that is,to distinguish the high-risk population and the low-risk population accurately.In this paper,based on mammography images,by mining the asymmetric features of the breast and combining them with a machine learning algorithm,we build a short-term breast cancer risk prediction model.The specific research contents and innovations of this paper are as follows:Firstly,the preprocessing of mammography images.In this study,the mammography image data preprocessing work,using Canny edge detection and boundary detection method to find out the minimum external rectangular region of the breast,good removal of the X-ray mammography background marker area,reduce the background interference factors.The method based on graphics and region growing is used to segment breast muscle to reduce the interference of the breast muscle region.Secondly,mining the asymmetric characteristics of the two breasts.The asymmetric features include asymmetric texture features and asymmetric deep features.Texture asymmetry features mainly extract three types of features: image texture features,statistical difference features and structural difference features.Then,through the selfdesigned depth network model,the deep-seated image semantic information is mined,and the depth asymmetry feature of the breast is extracted.After extracting the asymmetric features,the designed feature selection method is used for feature selection to extract the features that are helpful to the discrimination ability of the model.And in the experimental process,more image information is added through the feature fusion of two perspectives.The experimental results show that the fusion of two perspectives can indeed improve the accuracy and robustness of the model.Thirdly,recent breast cancer risk prediction based on the asymmetric characteristics of the two breasts.The model is constructed by extracting the asymmetric features of the two breasts and the machine learning algorithm.Compared with different machine learning algorithms,extreme gradient boost(XGBoost)was used as the basic model of this study to build a short-term breast cancer risk prediction model.In the experiment,the single view feature and dual view feature are used to build the model respectively,and the analysis and comparison show that the dual view feature fusion can improve the overall performance of the model.And the depth asymmetry feature is integrated into the model,which can further improve the performance of the model.By using the asymmetric feature and XGBoost classification algorithm to build the prediction model.The final model accuracy values were 0.712,precision values were0.708,recall values were 0.81,F1 values were 0.756,AUC values were 0.718.The model can play an effective role in guiding the prevention of high-risk groups of cancer,and has reference value. |
