Intelligent Diagnosis And Treatment Of Breast Diseases

Breast cancer has become one of the cancers with the highest incidence in the world,and it is also the cancer with the highest fatality rate among female patients.Previous studies have shown that the 5-year survival rate of early breast cancer is as high as 95%.Therefore,early detection and treatment are the most effective ways of the prevention and treatment of breast cancer.However,medical resources are scarce and unevenly distributed in China,and there is a lack of experienced doctors in some remote areas,which greatly impacts the detection and diagnosis of early breast lesions.Therefore,the study of intelligent diagnosis and treatment of breast diseases and the construction of corresponding computer-aided diagnosis systems can promote the sinking of medical resources and improve the efficiency and accuracy of breast disease screening.Deep neural network is a kind of computational model that simulates the information pro-cessing process of human brain neural networks.It has become an important method of artifi-cial intelligence and has achieved great success in many fields in recent years.It has achieved a performance comparable to that of human doctors in the intelligent diagnosis and treatment of breast diseases in some tasks.However,due to non-standard data collection,various types of indicators and discretization of key features in breast diagnosis and treatment data,the deep neural network method for breast intelligent diagnosis and treatment still faces many problems and challenges.Based on the real-world clinical requirements,in fronted with the difficulties in discriminant feature extraction,imbalance of sample size and high rate of missed diagno-sis in the process of intelligent diagnosis of breast diseases,this study focuses on the neural network method for intelligent diagnosis of malignant lesions in breast ultrasound images,in-telligent BI-RADS grading and missed diagnosis detection.In order to solve the problems of time-consuming,low consistency and multi-type constraints in the treatment of breast diseases,a breast cancer radiotherapy target volume delineation model based on dynamically strided con-volution neural network and a breast cancer radiotherapy dose prediction model based on dose intensity constraining are studied.Using the constructed neural network model,a number of computer-aided diagnosis systems for breast diseases have been developed and applied in real-world scenarios in many hospitals across the country,showing good clinical applicability.The main contributions of this study are listed below:1.Intelligent diagnosis and grading models of breast diseases are studied,feature enhance-ment and discriminative feature learning modules based on deep neural networks are pro-posed to realize accurate recognition and classification of breast lesions.A breast ultra-sound intelligent screening system and a real-time intelligent BI-RADS grading system have been developed,which have been applied in clinical practice in over one hundred of medical institutions in many provinces and cities across the country,and can effectively assist doctors in breast diseases diagnosis and improve work efficiency.Breast diseases are the inflammation,hyperplasia,tumor and other lesions that occur in the breast mass of patients.Ultrasound imaging is one of the important diagnostic meth-ods of breast diseases.Identification of breast malignant lesions and BI-RADS grading of breast diseases based on ultrasound images are important parts of extensive screen-ing of breast cancer and play important roles in the accurate diagnosis and treatment of breast cancer.In this study,the methods of intelligent breast diseases recognition and BI-RADS grading based on ultrasound images are studied,and the largest known breast ultrasound image data set is constructed.A region enhancement mechanism of deep neu-ral network feature extraction is proposed,through multi-class collaborative learning,the ability of extracting multi-dimensional features of breast lesions of the neural network can be enhanced,and the accuracy of breast malignant lesions recognition can be im-proved.In addition,this study proposes a discriminative feature learning method based on dynamic cost sensitive learning,which can dynamically adjust the learning weights of different samples,effectively alleviate the side-effects caused by the imbalance of multi-class data,and improve the accuracy of BI-RADS classification of breast diseases.Using the proposed model,a breast ultrasound intelligent screening system for individual users and a real-time intelligent BI-RADS grading system for ultrasound departments are con-structed to assist patients and doctors in the diagnosis of breast diseases,reduce the burden of doctors,and improve diagnosis accuracy and efficiency.2.Missed diagnosis detection model of breast ultrasound images is studied,and a recogni-tion method for missed diagnosis samples of intelligent diagnosis model is proposed by modeling breast ultrasound image data in two aspects: space consistency and distribution consistency.The proposed model can effectively reduce the missed diagnosis rate of the deep neural network intelligent diagnosis model.The ultrasonography diagnosis of breast cancer has direct influence on the way that doc-tors intervene patients.Patients diagnosed as malignant will take further examinations and treatment,while patients diagnosed as benign will be advised to go home and fol-low up later.Missed diagnosis may lead to disease deterioration and reduce the effect of treatment,and has become a serious social health problem,as it is widely spread in both manual diagnosis and intelligent diagnosis.In this study,a missed diagnosis detection method based on breast ultrasound image intelligent diagnosis model is studied.Based on the diagnosis results and intermediate features of the deep neural network diagnosis model,the space consistency of data is modeled by auto-encoders,and the distribution consistency of data is modeled by generative adversarial networks.The essential features of true negative samples are extracted.The false negative samples are regarded as abnor-mal samples for abnormal detection to realize accurate detection of false negative samples and reduce the missed diagnosis rate of the deep neural network diagnosis model.In this study,the proposed missed diagnosis detection model is evaluated on a variety of different types of medical image datasets,with the combination of many different architectures of deep neural network intelligent diagnosis models.The experimental results show that the method has good generalizability and is easy to expanse,and can be applied in a variety of medical image intelligent diagnosis tasks,effectively reducing the missed diagnosis rate and improving the sensitivity without a huge lost of diagnosis accuracy.In this case,the treatment effectiveness and survival rate of patients can be improved,and the side-effects caused by missed diagnosis can be alleviated.3.Intelligent breast cancer radiotherapy model is studied,a target clinical volume delin-eation model based on dynamically strided convolutional neural networks and a dose prediction model based on region distribution learning are proposed to realize accurate delineation of breast cancer radiotherapy clinical target volume and promising prediction of dose distribution map.Radiotherapy is an important treatment method for breast cancer,and has been widely used in preoperative preparation,postoperative cleaning and other scenarios.Radiother-apy planning is the key to ensure radiotherapy treatment effectiveness,target volumes delineation and dose prediction are important steps in the planning process.In the step of target volumes delineation,the clinical target volume area and the organs at risk are segmented accurately,and the detailed dose intensity distribution is generated according to the plan target volume area and the organs at risk contours in the dose prediction step.In this study,a target volume delineation method and a dose prediction method based on deep neural networks are proposed.In target volume delineation,a dynamically strided convolution operator and its corresponding dynamically strided fractionally convolution operation are proposed.Without introducing additional computation cost,the method can optimize the features extracted from the region of breast lesions,and effectively improve the accuracy of target volume segmentation.In addition,for the dose prediction task,a voxel loss function is proposed,which constrains the network training according to the dose distribution in the target volume area and the organs at risk.Besides,an order loss function is proposed to constrain the network training further,according to the orders of dose values in the regions.In practical clinical application,a patient’s radiotherapy plan usually requires 2 to 4 hours of target volume delineation and several days of dose prediction.The model constructed in this study can assist radiotherapists for accurate target volume delineation,shorten the cost time to a few minutes,and guide dose predic-tion optimization through the predicted dose map,reduce the number of iterations in dose prediction workflow,and improve work efficiency.