A Comparison Study Between The Simple Breast Lesions Classification Model Based On Texture Feature For X-ray Photography And Logistic Regression Analysis Scoring Model

Nowadays,mammary disease has become the most common disease in women,in addition to,breast cancer has become the most common female malignant tumor lesions.Breast lesions are divided into simple mass and complex lump.Simple mass refers only to bump as the main sign,which not included calcification or other symptom.Breast X-ray photography is the main means of disease diagnosis of breast calcification,the discriminant of the benign and malignant good diagnosis effect.But a large number of studies have shown that breast X-ray photography can't alone more accurate diagnosis of simple mass type lesions,because of its dense type of mammary gland tissue who's the Asian women had.Therefore,we collected 137 sample size for pure breast lesion cases,in combination with Clinical features?ultrasound and FFDM or DBTs' maging characteristics to build a simple scoring model.To analysis of simple mass lesions diagnostic value and preliminary discussed the simple breast lesions' benign and malignant of identification strategy.In recent years,the development of Radiomics and the program of Deep Learning,the study of diagnosis model in making diagnosis and giving treatment in lung cancer?radiation and chemotherapy after operation by the patient whom colorectal cancer.But,in fact,the study of mammary gland disease had seldom been completely.In this research,we try to combine the thought of Radiomics with the picture processing and classification in DBT image which is form breast mass lesion.Extracting and analyzing texture feature form the DBT image form breast mass lesion,use the meaningful feature to build up a mathematic model,then,training by stable SVM classifier.At last,using the classifier assisted radiologists to diagnose the mass after Deep Learning.This research including the follow four part:The first part:Introduction.In the part mainly expounds the background of the research,significance,makes a systemic analysis and research for the research situation of breast mass lesion and the preponderance or the disadvantaged about diagnose and classification.Using the common model,which included Mammography,Ultrasound and Clinical sign and symptoms,build up a simple score model.And then,describe the development of Radiomics,combined breast mass lesion with texture feature and Deep Learning,on account of them to build up a classification model.In order to make a preliminary assessment about the difference between two diagnose models'background and application prospect,provides a theoretical basis for the follow analysis.The second part:Build up a simply score model for breast simply mass diagnose and classification by Logistic regression mode.We collected 137 patients from Nanfang hospital at 2016.01-12,which were fit for the standard of breast simply mass lesion.We collected the date included Mammography(FFDM and DBT),Ultrasound and Clinical sign and symptoms,making a analysis by two elderly radiologists and picked up meaningful features.Combining them clinical experience and diagnosis with Logistic regression mode to make a assessment.Base on it,we could get the OR values and built up a simply score model to grade the mass.Making a correlation between the score and pathology to assessed the diagnostic efficiency about the model.Consequence:we got 90 cases benign and 47 cases malignant.Kappa score is 0.952.After summary,we collected 46 features,31 features would analyzed by Logistic regression mode,finally,we got 6 features into score model.Verge(OR=0.43),Substantial transformation(OR=6.99),Ultrasonic mass internal echo(OR=1.35),Age(OR=3.72),Stiffness(OR=0.05)and Mass activity(OR=0.87);The model's cut point is 0.5,when the score bigger than 0.5,we could identify it for malignant or when score less than 0.5,we could identify it for malignancy,AUC=0.8.The model sensibility is 87%,specificity is 70%,PPV(positive predictive value)is 60%,NPV(negative predictive value)is 91%.The value of Kappa means the goodness of fit with the two radiologists.All of the six features,protective factors are Verge,Stiffness and Mass activity;risk factors are Substantial transformation,Ultrasonic mass internal echo,Age.It means that the score model has better diagnostic efficiency to identify mass.The third part:Establishing simply mass classification model based on DBT image texture features extraction.In this part,we collected 137 patients whom from Nanfang hospital is fit for breast simply mass standard at 2016.01-12.Delineating the mass ROI by two elderly radiologists,and then,extracted the texture by Biomedical Engineering college.At last,we got 82 features,15 of them could extract for SVM model.Based on SVM classify model to set up study model and use it for crosscheck.As a result,we choose 13 textures to establish model by DBT-MLO image that we would get the best AUC,which is 0.76,and we could classify most of all mass into benign or malignancy.The fourth part:Contrastive analysis among Logistic regression mode simply score model,Texture extracted classification model and Clinical experience diagnose model.Finally,we collected 29 patients whom from Nanfang hospital is fit for breast simply mass standard at 2017.01-03.Then,using there three ways to identify each benign or malignant,build ROC curve one by one and AUC point.The purpose of this part is to acquire a more efficiency and standardization model.In consequence,it is texture extracted classification model has the same AUC point with clinical experience diagnose model,which is 0.82.Which has an equilibrium sensibility,is 0.78,specificity is 0.86.Texture model's high sensibility and slightly lower specificity means that is suitable for screening of large sample research.