The Value Of Radiomics And Deep Learning Based On Mammography For The Identification Of Clinical Risk And The Prediction Of Upgrade Of Ductal Carcinoma In Situ

BackgroundThe preferred imaging tool for ductal carcinoma in situ(DCIS)is mammography.Under the individualized treatment model,how to identify DCIS with different clinical risk levels based on mammography images has become an urgent problem in clinical work.In addition,the limitation of puncture biopsy sampling causes some patients with biopsy-proven DCIS to be found to have pathologically upgraded to infiltrative lesions during subsequent resection surgery,which poses a challenge to establish a non-invasive and efficient preoperative prediction model for pathologic upgrade of DCIS.Image histology and deep learning can quantify the extraction of different level features and become a research hotspot in the field of breast imaging,while the efficacy and accuracy of current mammography-based images using image histology to construct DCIS clinical risk level classification and pathology escalation prediction models are still insufficient.Objective1.To compare and analyze the clinical and X-ray performance of DCIS with different risk levels,to construct identification models of DCIS with different clinical risk levels based on Hand-crafted radiomics(HCR),Deep learning(DL)features and clinic image features and to explore their application value.2.To compare the clinical,pathological and X-ray manifestations of simple and pathologically upgraded DCIS and screen the predictors related to pathological upgrade,extract HCR and DL features and screen the optimal feature model,construct a joint model for predicting pathological upgrade of DCIS based on predictors and optimal features and explore its application value.Materials and methods1.Retrospectively,the study enrolled patients DCIS confirmed by postoperative pathology between January 2015 and November 2022 at Shenzhen Hospital of Southern Medical University,including the clinical,mammography imaging,and pathological data of 264 patients,which were randomly divided into 176 cases in the training set and 88 cases in the test set.The pathological histological nuclear grade of DCIS was divided into low risk group(low-moderate nuclear grade)and high risk group(high nuclear grade)according to their different clinical risk.Using t-test or Mann-Whitney U test,the χ2 test or Fisher’s exact probability test to compare the statistical differences in clinical and imaging data between DCIS groups with different risk.The semi-automatic detection and segmentation of MCs were completed by the self-developed program,and further the HCR features and DL features of MCs were extracted respectively.The feature selection was completed by one-way ANOVA,Pearson correlation test,and LASSO regression.A support vector machine(SVM)classifier was used to construct the risk classification model of DCIS,based on the selected HCR features,DL features and the combined Deep learning radiomics(DLR)features,clinical imaging features,clinical+HCR features,clinical+DL features,clinical+DL features,and clinical+DLR features,respectively.The efficiency of the model was evaluated by the receiver operating characteristic(ROC)curve with the area under the curve(AUC),accuracy,and specificity,to determine the most effective model.2.Retrospectively,336 patients with pathologically confirmed DCIS by puncture biopsy and subsequently undergoing excisional surgery at Shenzhen Hospital of Southern Medical University between January 2015 and November 2022 were retrospectively included and randomly divided into 252 cases in the training set and 84 cases in the test set.The patients with DCIS were divided into pure and pathologically upgraded groups according to whether pathological escalation to ductal carcinoma in situ with microinvasion(DCIS-MI)or invasive breast cancer(IBC)was found during the final surgery.The general clinical data,pathological and immunohistochemical results of the patients were recorded,and the mammographic manifestations,using ttest or Mann-Whitney U test,the χ2 test or Fisher’s exact probability test was used to compare the statistical differences of clinical,pathological and imaging data between the two groups.Single-and multi-factor logistic regression analysis were used to select independent predictors of clinical,pathological,and imaging features associated with upgraded DCIS.The semi-automatic detection and segmentation of microcalcifications were completed by a self-developed program,and further the HCR features and DL features of microcalcifications were extracted by imaging histology and deep learning techniques,respectively,and the feature screening was completed by one-way ANOVA,Pearson correlation test,and LASSO regression.The SVM classifier was used to construct prediction models for DCIS pathology escalation based on the screened HCR features,DL features and the combined DLR features of both,respectively,and the optimal machine learning model among them was evaluated and determined using the area under ROC.In addition,clinical-semantic prediction models were constructed based on independent predictors,combined models of clinicalsemantic and DLR features were constructed and as well as the nomogram.ROC curves and calibration curves were used to evaluate the predictive efficacy of each model for the upgrade of DCIS.Delong test was used to compare the differences in AUC values between models,and decision curve analysis was used to evaluate the clinical use value of the optimal efficacy models.Result1.In the training and test sets,lymph node abnormalities,BI-RADS classification,morphology and distribution of MCs were statistically different between the low-risk and high-risk DCIS groups(p<0.05).In the high-risk group,the following were more common:abnormal lymph node,with the rate of 3.7%(4/108)and 1.8%(1/57),respectively;BI-RADS 4A or higher classification results,with the rate of 65.7%(71/108)and 66.6%(38/57),respectively;and fine line or thread-like branching MCs,with the rate of 81.5%(88/108)and 73.7%(36/57);and linear or segmental distribution of MCs,with the rates of 63.9%(69/108)and 77.2%(44/57),respectively.And more common in the low-risk group:coarse inhomogeneous or indeterminate MCs,with the rates of 42.4%(29/68)and 41.9%(13/31),respectively;and cluster-like distribution,with the rates of 42.6%(29/68)and 54.8%(17/31),respectively.The feature extraction generated 544 HCR features and 256 DL features,and 36 high-value features were obtained after selecting.The recognition efficiency of the DL feature model in the test set(AUC=0.741,95%CI,0.697-0.795)was higher than that of the HCR feature model(AUC=0.657,95%CI,0.611-0.703),and the recognition efficiency of the DLR model after the combination of both was higher in the training set and test set(AUC=0.895,0.901)than that of HCR or DL models(P<0.001).The recognition performance of the combined model based on DLR features combined with clinical imaging features in the training set(AUC=0.913,0.932)(P<0.001)was 84.9%and 93.3%in accuracy and specificity,respectively,and the AUC was 0.932(95%Cl,0.893-0.979)in the test set,which were better than those of the single feature model,clinical imaging plus DL features and clinical image plus HCR features model(P<0.001).2.In both the training set and the test set,general clinical information was not statistically different in the comparison between groups for pure DCIS and upgraded DCIS.The included pathological factors including lesion diameter,ER,PR,HER-2,Ki-67,comedonecrosis,and histological nuclear grade were statistically different in the comparison between the two groups,and patients with pathologically upgraded DCIS in the training set and test were more likely to have larger lesion diameters(median 21 mm,23 mm,respectively),positive HER-2(occurrence rates of 61.4%,62.7%),positive Ki-67(51.3%,40.7%,respectively),comedonecrosis(72.7%,86.4%,respectively),and high nuclear grade(70.3%,69.4%,respectively),while patients with pure DCIS were more likely to have smaller lesion diameters(median 14mm,15mm,respectively),positive ER(94.7%,92.7%,respectively),positive PR(92.6%,88.0%,respectively),and low-moderate nuclear grade(63.8%,68.0%,respectively).The imaging factors included breast background,,and morphology and distribution of MCs were statistically different in the two groups,with fine polymorphic or linear branching microcalcifications(63.9%and 66.1%occurrence,respectively)and linear or segmental MCs(74.0%and 76.3%occurrence,respectively)more common in patients with upgraded DCIS and pure DCIS patients were more likely to have smaller coarse inhomogeneous or indeterminate(64.9%,68.0%,respectively)and clustered distribution of MCs(60.6%,52.0%,respectively).In univariate logistic regression analysis,lesion diameter of 2 cm,fine polymorphic or linear branching MCs,linear or segmental MCs,Ki-67 of 14%,high nuclear grade,and comedonecrosis were positively associated with upgrade of DCIS(P<0.05),while coarse inhomogeneous or indeterminate,clustered MCs were negatively associated with them(P<0.05)..In the multifactorial logistic regression analysis,only Ki-67 ≥14%(OR=0.04,95%CI:0.0003-0.44;P=0.005)and high nuclear grade(OR=0.44,95%CI:0.28-0.68;P<0.001)were independent predictors of upgrade of DCIS.The predictive efficacy AUC of the DL model was 0.743(95%CI,0.701-0.807)higher than that of the HCR model in the test set.Using 544 HCR features extracted by radiomics and 362 DL features extracted by deep learning,48 high-value feature sets were obtained after selecting,and the AUC of the DLR model after combining the two was 0.899 and 0.905 in the test set and training set,respectively,which were higher than those of the HCR and DL models(P<0.001).The AUC of the combined model with clinicalsemantic and DLR features improved by 3.2%and 21.4%(0.928 vs.0.899,0.928 vs.0.764)in the training set and 2.8%and 27.3%(0.931 vs.0.905,0.931 vs.0.731)in the test set,respectively,compared with the DLR and clinical-semantic models alone.The predictive efficacy of the joint model was higher than that of the clinical-semantic model and DLR model in the prediction of upgrade of DCIS(P<0.001).DCA showed that the joint model had higher clinical value than the clinical-imaging,DLR model,and column plots showed that DLR features had the greatest predictive weight in the joint model.Conclusion1.The model constructed by combining DLR features with clinical imaging features by imaging histology and deep learning can effectively identify different risk levels of DCIS,and the identification efficacy is higher than that of other single-class features or two mixed feature models,which can provide a new approach and reference to solve the problem of preoperative noninvasive evaluation of clinical risk levels of DCIS.2.The DLR feature model combining imaging histology and deep learning has higher prediction efficacy for pathological escalation of DCIS than single-class feature models,and the joint model constructed by adding clinical independent predictors to DLR features obtains the best recognition efficacy,which is higher than the clinicalsemantic,DLR feature model,and can provide new approaches and new references to solve the problem of predicting pathological escalation of DCIS to avoid overtreatment and improve prognosis.reference.