Neural Network Approaches For Intelligent Diagnosis And Treatment Of Breast Cancer

Breast cancer is the malignant tumor with the highest incidence rate and mortality.Statistics show that the five-year survival rate of breast cancer has reached approximately 95%,and early diagnosis and treatment can sharply reduce mortality.Mammography is the most extensive and effective for early breast cancer screening.Radiation therapy,which can improve patients’ living quality and prolong survival,is an effective treatment for breast cancer.However,the disease may not be detected and treated in time due to the irrational allocation of medical resources,the lack of high-quality medical resources,and inexperienced doctors in some remote areas.Thus,studying artificial intelligence methods in diagnosing and treating breast cancer has real meaning.Synchronously,building computer-aided systems can supply clinical references,which can alleviate doctors’ workload and elevate diagnostic efficiency and accuracy.Deep neural network methods have succeeded in many fields and permeated through medical fields,especially medical image analysis.However,breast medical images are significantly different from natural images.It embodies the complexity of the object relationship,more information,smaller focus areas,and more finely discriminative characteristics.Therefore,neural networks still face many difficulties and challenges in the diagnosis and treatment of breast cancer.Based on clinical requirements,aiming at the current problems,this paper studies the diagnosis of single-view digital mammograms and multi-view films mammographic film images.Besides,according to the defects in quality control of breast cancer volumetric intensity modulated arc therapy(VMAT)plans,this paper carries out the research on quality assurance prediction model of breast cancer VMAT plans and builds the intelligent auxiliary diagnosis and treatment systems for breast cancer.The main innovations of this thesis are as follows:1.Classification model of mammograms with multiple perspectives regional enhancingA region-enhanced multiple perspectives input method and a weighted multi-instance learning method are proposed to solve the challenges of large image size,the small proportion of lesion area,the slight difference between lesions and breast tissue,and imbalanced data.For the multiple perspectives input method,the gamma correction algorithm enhances the different regions in mammograms to enlarge the morphological and texture features of the lesions from the surrounding breast tissues.Then,this method integrates the original and the enhanced images as input of networks,which can retain the feature in the original image and obtain the more discriminative features in enhanced images.In the weighted multi-instance learning method,the mammograms can be divided into many regions by the global feature map,and the malignant probability of each region is calculated separately.Based on these probabilities,the most likely malignant region in the image is selected to present the properties of the entire mammogram.During the training process,designing a weighted loss function to reweight categories can effectively address the propensity of prediction and improve model performance.Experimental results show that the proposed model can effectively improve the diagnostic accuracy of mammograms and can roughly locate the lesions.2.Classification model based on multi-view mammograms with weight sharingDue to the limitations of the previously proposed single-view mammogram classification method in multi-view mammograms breast cancer diagnosis,this work proposed a classification model based on multi-view mammograms with weight sharing,which was more suitable for clinical actual scene diagnosis of breast cancer.Aiming at the problems of massive parameters,structural symmetry,and regional overlap between different view mammograms,selection of crucial features,and multi-view feature fusion,this work proposes a model that combines a multi-branch shared network structure and a top-k extreme value fusion method.Different view weight sharing and global weight sharing strategies are introduced to different branching networks,which extract image features from different views in the multi-branch network architecture.The top-k extreme value fusion method includes global top-k and local top-k extreme value selection two techniques.The local extreme value strategy filters the characteristics of suspected lesion areas in each view,and the global extreme value strategy selects the characteristics of suspected lesion areas in the entire case.The features filtered by the two strategies are fused to make up a discriminative feature set of multi-view mammograms for final prediction.In the experiments,the classification model uses the multi-view mammogram dataset,and the results show that the proposed model can exceed existing methods in terms of evaluation indicators.Based on the proposed model,an intelligent multi-view mammogram diagnosis system is constructed to assist the clinician in diagnosing breast cancer,and the prediction results and lesion location are used to assist doctors in rapid diagnosis of mammography images.The system has been put into clinical trial,and the practical application verifies that the proposed method can reduce the misdiagnosis rate and improve the diagnosis efficiency.3.Multi-modal quality assurance prediction model with symmetric fusion on different data for breast cancerThe breast cancer quality assurance(QA)prediction models of VMAT plans have many difficulties,including the lack of key information,differences in multi-modal data di-mension,and multi-modal fusion.According to these difficulties,this work proposes a multi-modal breast cancer QA prediction model with symmetric fusion on different data,which integrates an equal dimensional feature mapping and symmetric fusion with different data module.The former extracts the image modal features and maps the highdimensional image features to a low-dimensional feature space with the same dimension as the non-image modality.It can reduce the dimension discrepancy between two modalities.The latter utilizes a symmetric structure to fuse two types of data sources,including the features of image modality and the source data of non-image modality.Then,it learns complementary information from the fused features to predict the gamma passing rate(GPR)value of the radiotherapy plan.In this study,a multi-modal breast cancer QA dataset and a multi-disease cancer QA dataset are constructed,and a large number of experiments are carried out on these two datasets to verify the effectiveness of the proposed model.Experimental results show that the proposed model can achieve good performance in QA prediction tasks.In addition,an intelligent quality control verification system for VMAT radiotherapy plans is constructed based on the proposed model,and tested in the radiology department of West China Hospital.It helps radiologists quickly obtain the QA results of VMAT plans,evaluate the quality of VMAT plans,and improve the efficiency of radiotherapy QA workflow.4.Multi-modal neural networks based on association mapping for quality assurance of breast cancerOn the basis of the multi-modal prediction model proposed earlier,there are still many problems in the multi-modal QA prediction model of breast cancer,consisting of the feature extraction of different types of images,association learning within multi-modal data,and significant difference between multi-modal data.To solve these problems,this work further presents a multi-modal breast cancer QA prediction model based on association mapping.It mainly contains the association mapping network(AMN)module and local alignment-based fusion(ABF)module.The former learns the relationship between twodimensional and three-dimensional images.In AMN module,a multi-branch network is designed to extract the 2D and 3D image features in the image modality respectively,and then the local prediction results of the whole image modality are calculated through the two types of image features.It converts the high-dimensional image features into the low-dimensional representation of non-image modality,thereby reducing dimensional differences and realizing transformation between two modalities.The ABF module fuses the local results of image modality with the non-image modality through the consistency of modalities,and learns the corresponding relationship between the two modalities to predict the GPR.In order to verify the effectiveness and validity of the proposed model,many experiments are executed on two QA datasets.The results indicate that the proposed model can accurately predict the GPR value of breast cancer VMAT plans and perform better than other multi-modal QA methods.