The Study And Application Of Breast Tumor Classification Based On Casual Stability Learning

Breast cancer is most prevalent in women,and its incidence ranks first among malignant tumors in women and is the second most important factor in female mortality,posing a serious threat to women’s health.Clinical studies have so far failed to identify the direct causative factors of breast cancer,which makes there is no effective primary prevention measure for breast cancer,therefore,early diagnosis is especially important.Among breast cancer examinations,mammography has become the main adjuvant examination for breast tumors because it is more sensitive to microcalcifications,and its examination report is an important basis for clinicians to determine the nature of breast tumors.Therefore,using mammography reports to assist doctors in breast tumor diagnosis and alleviate the pressure on doctors in breast tumor screening is an important manifestation of smart medicine and intelligent treatment.At present,deep learning models for breast tumor diagnosis have achieved good results,but the model stability is poor because they ignore the complex selection process of actual data,i.e.,the possible distribution bias problem of model testing and training data.And most of these models lack interpretability,while the diagnostic and prognostic models in medical field are extremely sensitive to the safety and reliability of the models,and the inability to provide doctors with sufficiently reliable interpretation of decision results is a major reason why the models cannot be popularized in actual clinical practice.To achieve accurate diagnosis of benign and malignant breast tumors based on mammogram reports and to improve the stability and interpretability of the model,this paper proposes a breast tumor classification model（SSC-Tab Net）based on causally stable learning with real data from a tertiary hospital in Shanghai,then combines the self-explanatory and post-hoc interpretable methods of the model to provide interpretable analysis of the decision results.In order to assist clinicians in diagnosis,this paper designs and builds a breast tumor assisted diagnosis system based on this model.The research in this paper is divided into the following three parts.（1）Construction of a breast tumor diagnosis model based on casual stability learningIn this part,an SSC-Tab Net model is proposed.Firstly,the data set is prepared based on the composition and structural characteristics of mammogram reports.Then,the mammogram report is structured according to the "segment-tissue descriptor-attribute descriptor" rule based on the semantic tree of the mammogram to obtain a rich semantic hierarchy of the mammogram report.After that,the input data is mapped to a low-dimensional space using nonlinear functions to reduce the feature dimensionality while preserving the original feature information,thus alleviating the missing data situation.At the same time,being in the low-dimensional space helps the subsequent learning of global balance weights.Finally,the global balance weights are learned using causal inference,and the weights are used to guide the feature selection process of the model,which is used to improve the stability of the model.The experimental results show that the proposed model in this paper has better accuracy and stability than the traditional classification algorithm for tabular data.（2）Realization of interpretability analysis based on model self-explanation and model post-hoc explanationThis section provides an interpretability analysis of the SSC-Tab Net model,first analyzing the internal mechanism of the model for deciding breast tumor properties,calculating the weighted sum of the weight coefficient matrix for each decision step as the importance score of breast tumor features,and achieving self-interpretation of the model.Then,the post hoc interpretability analysis of the model was used to obtain the contribution of any breast tumor feature to the model decision,which provides a complementary interpretability analysis for the model.In this paper,we combine model self-interpretation and model post-hoc interpretation methods to jointly provide interpretable analysis of the model and improve the reliability of the model.（3）Design and construction of a breast tumor assisted diagnosis systemIn order to apply the model to clinical practice and effectively assist doctors in diagnosis,this part designs and builds a breast tumor assisted diagnosis system.This part analyzes the user requirements,functional requirements and use case design from the doctor’s perspective,and completes the overall framework of the system,the technical scheme of implementation and the design of database based on it.Finally,it realizes the functions of mammography report management,diagnosis and interpretable analysis of the report.