Using Machine Learning Method To Integrate Ultrasound Multi-parameters To Construct A Prediction Model Of Breast Intraductal Carcinoma In Situ

Background: Breast cancer is a kind of common tumor in Chinese female population.The 2020 WHO survey shows that the global incidence of breast cancer is 17.1%,and the mortality rate is also high,reaching 8.16%.The number of breast cancer patients in the world has exceeded that of lung cancer,and the number of patients ranks first among all kinds of cancers.At present,ductal carcinoma in situ(Ductal carcinoma in situ,DCIS)has become a common type of breast cancer,accounting for 15-25%.With the upgrading of ultrasound machines,the diagnosis rate of intraductal tumors is significantly higher than before.because the clinical and morphological features of intraductal papilloma and intraductal carcinoma in situ are very similar,it is difficult to distinguish between intraductal papilloma and intraductal carcinoma in situ.Some studies have shown that the combination of multi-parameter analysis can significantly improve the diagnostic accuracy.Therefore,the use of imaging to predict breast ductal carcinoma in situ has a certain research value.Based on this,our hospital specially designed this experiment.Objective: The purpose of this study is to use machine learning-related ultrasound parameters to find out the imaging features of predicting intraductal carcinoma in situ and to construct a differential diagnosis model of intraductal lesions,so as to screen out intraductal lesions in situ early and reduce invasive examinations such as puncture.Improve the accuracy of treatment.Methods: This study retrospectively collected the clinical information of 289 female patients with breast intraductal lesions confirmed by pathology in our hospital from March 2019 to July 2022.256 patients were taken as the training set,their ultrasonic features were extracted,and the differential model of benign and malignant intraductal lesions was constructed according to the pathological results,and 33 patients were taken as the verification set to verify the accuracy of the model constructed in this study.LOGIQE9/E8 and other ultrasonic diagnostic instruments were used to examine the hierarchical structure of the patient’s breast.The age(age),maximum diameter of the tumor(mm),regular shape(0,no 1),smooth edge(smooth 0,rough 1),ultrasonic classification(1),aspect ratio(≥ 1)and aspect ratio(≥ 1)were recorded.The ultrasonographic and pathological features were as follows: microcalcification(no 0,single 1,multiple 2),distance from nipple(mm),posterior echo(constant 0,enhanced 1,attenuation 2),internal blood flow signal of the tumor(no 0,a small amount of 1,medium 2,a large number of 3),peripheral blood flow signal(no 0,a small amount of 1,medium 2,a large number of 3),tumor type(central type 0,peripheral type 1)and cancer(yes 1,no 0).According to the pathological results,the relevant ultrasonic features of intraductal papillary carcinoma and intraductal papilloma can be extracted from the training group.After statistical processing,the variables of P < 0.10 were included in binary logistic regression analysis to build a machine learning model for differential diagnosis of breast intraductal papilloma in situ and intraductal papilloma.Finally,the accuracy and diagnostic efficiency of the model are verified by the verification group.Results: According to the gold standard of postoperative pathology,the statistically significant features for differentiating intraductal papillary carcinoma from intraductal papilloma included tumor maximum diameter,smooth edge,ultrasonic classification,microcalcification,internal blood flow signal and peri-tumor blood flow signal(P < 0.05).The maximum diameter of the tumor,whether the edge was smooth,ultrasonic classification,microcalcification,blood flow signal within the tumor,blood flow signal around the tumor,age,regular shape and the distance between the lesion and the nipple were included in binary logistic regression analysis(condition: LR.(forward)it is found that the model with the best diagnostic effectiveness is the machine learning model for differential diagnosis of intraductal lesions based on whether the edge is smooth,ultrasonic classification,microcalcification,peri-tumor blood flow signals and age: Model formula: differential diagnosis of intraductal lesions = B edge * & + B calcification * & + B peripheral blood flow signal * & + B age * & + B ultrasonic typing * &(B is the regression coefficient of the model corresponding to variable.& the value is 0 or 1,and 1 when the variable is positive,otherwise it is 0).The area under the working characteristic curve(Area Under Curve,AUC)of the subject working characteristic curve(receiver operating characteristic curve,ROC)curve for distinguishing intraductal carcinoma in situ from intraductal papilloma in the training group was 0.776(95% CI: 0.721 ~ 0.832).Among the 33 patients with intraductal lesions,there were 14 cases of intraductal papillary carcinoma and 19 cases of intraductal papilloma,with an average age of(44.36 ±9.92)years.Taking postoperative pathology as the gold standard,the AUC of the differential diagnosis model of intraductal lesions in the verification group to distinguish intraductal carcinoma in situ from intraductal papilloma was 0.914(95% CI: 0.821 ~ 1.000).Conclusion: In this study,a differential diagnosis model of intraductal lesions is constructed based on machine learning-related ultrasound and pathological features,which is expected to improve the accuracy of differential diagnosis of intraductal lesions and help to screen early carcinoma in situ in intraductal lesions.Reduce puncture and other invasive examinations to improve the accuracy of treatment.