| Objective: To investigate and compare the value of radiomics model,clinical model and combined omics-clinical model based on multi-parameter magnetic resonance imaging in the differential diagnosis of benign and malignant breast lesions,and to conduct BI-RADS(Breast Imaging Reporting Reporting)based on the output results of the dominant model.and Data System)quantitative scores to provide a feasible basis for clinical decision-making.Methods: The medical records of patients who underwent breast MRI examination in our hospital from January 2016 to January 2020 and underwent surgery or needle biopsy within two weeks after the examination were retrospectively collected,and 132 standard patients with 134 lesions were finally included.Due to the uneven distribution between benign lesions(n=43)and malignant lesions(n=91),the samples were first balanced with the Synthetic minority oversampling technique(SMOTE),and then the included All lesions were randomly divided into training and test groups in a ratio of 7:3.On the one hand,the ROI was delineated by 3Dslicer(version 4.13.0),and the image features of each sequence of lesions were extracted by the package Pyradiomics in the open source software python(version 3.9.7)to obtain five datasets,namely Tirm dataset and ADC dataset.,Enhance the first-phase data set,enhance the third-phase data set and the delayed-phase data set,and gradually use ANOVA,univariate analysis,random forest,Least absolute shrinkage and selection operator(LASSO)to carry out Dimensionality reduction,the model score is calculated according to the weighting coefficient of each variable,and the single sequence omics scoring model and the combined sequence omics scoring model are established in turn.On the other hand,univariate analysis was performed on clinical features and general imaging features in the training group,followed by multivariate Logistic regression analysis,and variables with P<0.05 were screened to construct a clinical model,and finally a Nomogram model of the omics-clinical combination was constructed.Receiver operating characteristic curve(ROC)and Hosmer-Lemeshow test were used to analyze the discrimination and calibration of each model in the two data sets,and calculate the area under the curve(AUC)of ROC,Sensitivity,specificity,and diagnostic accuracy.Differences in AUC were compared between combined sequence omics score models,clinical models,and combined omics-clinical models using Delong’s test.Decision curve analysis(DCA)was used to evaluate the net benefit of the nomogram model.Finally,the BI-RADS quantitative score is carried out with the output of the dominance model to guide clinical decision-making.Results: In the training group single-sequence omics scoring model,the AUC value of the omics scoring model based on the delay period was the largest,which was 0.952(95% confidence interval:0.900-0.982),and the maximum AUC value based on Delayed＿ADC＿dyna VIEW3 in the omics scoring model of the combined sequence was 0.964(95% confidence interval: 0.915-0.989),but after the Delong test,it was not statistically different from the Delayed＿ADC-based model(P=0.372).Multivariate logistic regression analysis of clinical and general imaging features showed that age(OR=1.111,95% confidence interval was 1.020-1.211,P=0.016)and spur(OR=9.210,95% confidence interval was 4.916-17.261,P<0.016)0.001)had a statistical difference between groups,so as to construct a clinical model,the obtained AUC value was 0.919(95% confidence interval: 0.857-0.960).Incorporating the omics score with age and spur,a combined omics-clinical model was constructed,and the obtained AUC was 0.986(95%confidence interval: 0.947-0.999).Comparing the three models,it can be seen that the combined omics-clinical model had the largest AUC value,and after the Delong test,it was statistically different from the models based on clinical-only and omics-only scores(P values were 0.001 and 0.036,respectively).The AUC values of the omics scoring model,clinical model,and combined omics-clinical model in the test group were lower than those in the training group,which were 0.735(95% confidence interval:0.599-0.845),0.887(95% confidence interval: 0.773-0.957),0.890(95% confidence interval: 0.776-0.958),respectively.The Hosmer-Lemeshow test indicated that the combined omics-clinical model fit was better for both the training and test groups,and the training group was better.The results of decision curve analysis show that the joint model in the training group has a higher net benefit,and the positive net benefit can be obtained in the test group when the threshold probability is less than 0.8,but it is slightly lower than that in the training group,and the threshold probability is 0.8-0.9.Get a negative net gain.The predicted probability p was classified in the training group:(1)BI-RADS category 3: 0<P≤2%,34 cases,all benign;(2)BI-RADS category 4a: 2%<P ≤ 10%,12 cases,of which 1 case was malignant and 11 cases were benign;(3)BI-RADS class 4b: 10%<P≤50%,12 cases,of which 1 case was malignant and 11 cases were benign;(4)BI-RADS class 4c: 50%<P<95%,19 cases,of which 15 cases were malignant and 4 cases were benign;(5)BI-RADS category 5: 95%≤P<100%,50 cases were malignant.The predicted probability p was classified in the test group:(1)BI-RADS class 3: 0 < P ≤ 2%,18 cases,including 1 malignant and17 benign;(2)BI-RADS class 4a: 2% < P ≤ 10%,5 cases were benign;(3)BI-RADS category 4b: 10%<P≤50%,11 cases,of which 7 were malignant and 4 were benign;(4)BI-RADS category 4c: 50%<P<95%,7 cases,including 3 malignant and 4 benign;(5)BI-RADS category 5: 95%≤P<100%,14 cases,including 12 malignant and 2 benign.Conclusion: In the classification and identification of benign and malignant breast lesions,the combined omics-clinical model established by the combined omics score of Delayed＿ADC and clinical factors(age,burr)is superior to the model constructed only by the omics score and the model constructed only by clinical.The omics-clinical combined model has good diagnostic performance in both the training group and the test group.After BI-RADS quantification with the dominant model,it can more clearly provide a feasible basis for the accurate classification of breast lesions,so as to provide patients with better results.for better clinical decision-making. |
PDF Full Text Request