| Objective: To check out the significance of radiomics features abstracted from contrast enhanced T1-weighted imaging(T1+C)in discriminating triple negative invasive breast cancer from non-triple negative invasive breast cancer,so as to provide non-invasive imaging predictive markers for clinical treatment decisions.Methods: 199 patients who underwent MRI examination in our hospital from January 2018 to December 2020,and finally diagnosed with invasive breast cancer were retrospectively enrolled in this study.The clinicopathological data and the general imaging features such as lesion morphology,enhancement characteristics,and the type of time-signal intensity curve were analyzed and summarized.Patients were divided into the training set(138 cases)and the test set(61 cases)in a ratio of 7:3,then the differences of clinicopathological data between the training set and the test set were compared.Univariate logistic regression analysis was performed on clinicopathological features and general imaging features in the training set,and variables with P < 0.1 were selected for subsequent analysis.Volumes of interests were delineated by two physicians with the dedicated radiomics analysis software A.K.,and the radiomics features of the dynamic enhanced images in the early phase were extracted.The inter-observer consistency was analyzed by intraclass correlation coefficient(ICC).Then features with ICC >0.75 were retained.Variance threshold,spearman correlation analysis,minimum-redundancy maximum-relevance and least absolute shrinkage and selection operator logistic regression were used in the training set to screen features.Generalized linear regression models based on radiomics features and clinical factors were established respectively.Backward stepwise logistic regression analysis was employed in the models,and variables were further screened according to the Akaike information criterion.Scores of models were calculated according to the weighted coefficient of each variable in the models.The combined prediction model was established by the radiomics score and clinical factors selected by univariate analysis,and the corresponding score was calculated.Differences of scores between groups were compared by independent samples t test or Mann-Whitney U test.Receiver operating characteristic curve(ROC)and Hosmer-Lemeshow test were used to analyze the differentiation and calibration ability of models in the two data sets,and the area under the curve(AUC),sensitivity,specificity,diagnostic accuracy,positive predictive value(PPV)and negative predictive value(NPV)of the ROC curve were calculated.Delong test was utilized to compare AUC of different models.Decision Curve Analysis(DCA)was used to assess the net benefit of the models to guide clinical decision-making.Results: 1.The median age of 199 patients was 49 years old,ranging from19 to 75.87 patients were postmenopausal,and 112 patients were premenopausal.Results of molecular subtype displayed 21 cases of Luminal A,77 cases of Luminal B,54 cases of HER2 overexpressing,and 47 cases of triple negative.The training set was composed of 32 cases of triple negative patients and 106 cases of non-triple negative patients,while the test set was composed of 15 cases of triple negative patients and 46 cases of non-triple negative patients.There were no statistically significant differences in the clinical data(age,menstrual status,immunohistochemical markers status)between the training set and the test set.Univariate logistic regression analysis manifested that there were statistically significant differences in age(OR=0.484,95%CI:0.213-1.108,P=0.083),Ki-67 expression status(OR=4.765,95%CI:1.553-20.853,P=0.015)and internal enhancement characteristics among groups(heterogeneous enhancement: OR=3.795,95%CI:0.705-70.54,P=0.209/annular enhancement: OR=18.75,95%CI:2.846-377.238,P=0.01).2.A total of1316 radiomics features were abstracted from each patient in the training set,and finally one original texture feature,two first-order features transformed from wavelet,one texture feature transformed from wavelet and one texture feature transformed from laplace of gaussian filter were obtained through feature selection to build the radiomics model.Three variables including age(OR=0.482,95%CI:0.198-1.18,P=0.107),Ki-67 expression status(OR=4.055,95%CI:1.238-18.552,P=0.037),and internal enhancement characteristics of lesions(heterogeneous enhancement: OR=2.729,95%CI:0.455-52.502,P=0.36/annular enhancement: OR=12.97,95%CI:1.768-271.956,P=0.029)were included in the clinical prediction model.Three variables including radiomics score(OR=2.879,95%CI:1.757-5.181,P<0.001),age(OR=0.408,95%CI: 0.153-1.069,P=0.068)and Ki-67 expression status(OR=8.144,95%CI:2.124-54.164,P=0.008)were included in the combined prediction model.There were statistical differences in clinical score,radiomics score and combined score between groups both in the training set and the test set.In the training set,the AUC,sensitivity,specificity and accuracy of the radiomics model were 0.762(95%CI:0.666-0.858),71.9%,73.6%,73.2%,respectively.The AUC,sensitivity,specificity and accuracy of the clinical prediction model were 0.747(95%CI: 0.656-0.837),59.4%,79.2%,74.6%,respectively.The AUC,sensitivity,specificity and accuracy of the combined prediction model were 0.830(95% CI: 0.747-0.914),71.9%,84.9%,81.9%,respectively.In the test set,the AUC,sensitivity,specificity and accuracy of the radiomics model were 0.743(95%CI:0.593-0.894),60.0%,65.2%,63.9%,respectively.The AUC,sensitivity,specificity and accuracy of the clinical prediction model were 0.696(95%CI:0.535-0.857),53.3%,67.4%,63.9%,respectively.The AUC,sensitivity,specificity and accuracy of the combined prediction model were 0.755(95% CI: 0.610-0.900),60.0%,73.9%,70.5%,respectively.Comparison of AUC between models in the training set showed statistically significant difference only between the combined prediction model and the clinical prediction model(P=0.015),while the P value of comparison between the combined prediction model and the radiomics model was 0.051,which nearly reached the significance level.There were no statistically significant differences in AUC among models in the test set.Hosmer-Lemeshow test showed that all models in the training set and the test set had good fitness(P>0.05).The results of decision curve analysis indicated that in the training set the overall clinical net benefit of the combined prediction model was higher within a large range of threshold probability,while in the test set the net benefit of the combined prediction model and the radiomics model was similar when the threshold probability is low,and the net benefit of the radiomics model was slightly higher than that of the combined prediction model when threshold probability was high.Conclusion: The combined prediction model consists of T1+C radiomics score and clinical factors including Ki-67 and age can be used for the differential diagnosis of triple negative and non-triple negative invasive breast cancer,and it has good diagnostic efficacy in both the training set and the test set.Radiomics model and combined prediction model both have certain clinically practical value,suggesting that using radiomics features combined with clinical factors to non-invasively predict the molecular subtypes of invasive breast cancer is feasible.However,whether the combined prediction model is more valuable than the radiomics model needs further exploration by increasing the sample size and regulating the research conditions. |
PDF Full Text Request