Value Of Ultrasound Radiomics Combined With Clinical-ultrasound Data In Predicting Lymphovascular Invasion Of Breast Cancer

Objective By constructing an ultrasound radiomics model and analyzing the significant features in the clinical-ultrasound data of patients with breast cancer,the value of the comprehensive model in the diagnosis of breast cancer lymphovascular invasion is explored,aiming at providing the reference information for clinicians to propose personalized diagnosis and treatment plan of breast cancer.Method This study retrospectively collected 150 female patients diagnosed with breast cancer by intraoperative or postoperative pathological results who underwent breast ultrasound examination and were hospitalized in our hospital from August 2018 to August2022.Clinical-ultrasound data and complete standard 2D ultrasound images(maximum cross-diameter section of lesions)were collected from all enrolled patients.According to the pathologic findings,the patients were divided into experimental group(45 cases of lymphovascular invasion)and control group(105 cases of non-lymphovascular invasion),and the features with statistical significance(P<0.05)in clinical-ultrasound data were screened.By manually delineating the region of interest(ROI)of the lesion in the 2D ultrasound images stored in DICOM format and extracting the ultrasound radiomics features,then the patient was randomly divided into training dataset and test dataset according to the ratio of 7:3,and the least absolute shrinkage and selection operator(LASSO)was adopted to screen the ultrasound radiomics features,finally used Logistic regression to construct the ultrasound radiomics model,then combined with significant clinical-ultrasound features to obtain a comprehensive model.Draw receiver operating characteristics curve(ROC),and compare the areas under curve(AUC),sensitivity and specificity of models respectively.Delong test and Hosmer-Lemeshow test were used to evaluate the performance of each model.Decision curve analysis(DCA)was used to compare the predictive value and clinical benefit of three models on breast cancer lymphovascular invasion.Statistical analysis Excel 2010 software was used for data entry,SPSS26.0 software was used for statistical analysis,measurement data were expressed as mean±standard deviation((?)±s),counting data were expressed as rate(%).Clinical-ultrasound data were compared by single factor binary logistic regression.The test level was α=0.05,and P<0.05 was considered to be statistically significant.As for ultrasound radiomics analysis,the delineation of the region of interest is realized in 3D Slicer 5.1.0 software,feature extraction is made using the open-source software package Pyradiomics platform in 3D Slicer5.1.0 software,the reduction of feature dimensionality and the construction of models is completed by Python software.Results1.Single factor and multivariate binary Logistic analysis of clinicalultrasound data in the experimental group and control group showed that axillary lymph node metastasis status and left/right sided breast were independent risk factors for breast cancer lymphovascular invasion(P<0.05).There were no significant differences in others(P>0.05).2.In the training dataset and the test dataset,the ROC of axillary lymph node metastasis status was drawn respectively to predict LVI,the AUC were 0.634 and 0.659,the ROC of left/right sided breast was drawn respectively to predict LVI,the AUC were 0.645 and 0.595,indicating that the accuracy of prediction alone is not high.Both of them predicted better,the AUC were 0.705 and 0.719.3.837 features were extracted from the ROI of each 2D ultrasound image,and 5 significant ultrasound radiomics features were finally acquired after screening the extracted features,the AUC of ultrasound radiomics model for predicting LVI were 0.820 and 0.833 in the test dataset and training dataset respectively.4.The 5 ultrasound radiomics features combined with axillary lymph node metastasis status and left/right sided breast were synthesized into a comprehensive model,and the AUC of ROC in the test dataset and training dataset were 0.899 and 0.844 respectively,which were higher than the two alone to predict LVI.5.Delong test showed that there was no significant difference in clinical predictive performance between the ultrasound radiomics model and comprehensive model.Hosmer-Lemeshow test showed that the above two models did not deviate from the fitting.Through DCA outcomes,the use of the two models to assist clinicians with making intervention decision is the best,especially,the comprehensive model has the best effect.Conclusion The comprehensive model constructed by ultrasound radiomics features combined with clinical-ultrasound data can better predict lymphovascular invasion of breast cancer,preoperative individualized noninvasive prediction of LVI in breast cancer patients brings reference information for more accurate personalized treatment of breast cancer patients.