| Objectives:There are some major changes in the latest fifth edition of the ultrasound(US)Breast Imaging Reporting and Data System(BI-RADS),including changes and additions to the description of lesion terminology,and incorporation elasticity assessment as one of the associated features to assess the risk of malignancy of breast lesions.The purpose of this study was to evaluate the application of the fifth edition of US BI-RADS in clinical practice and to investigate the diagnostic value of BI-RADS incorporated with strain elastography(SE)in the diagnosis of benign and malignant breast lesions by reclassifying BI-RADS category 3 and 4a lesions.Methods:A prospective,multicenter study was conducted in 32 hospitals in China,enrolling a total of 5012 patients.All patients underwent US and SE examinations to record US characteristics,associated features(shape,orientation,margins,echo pattern,posterior acoustic features,calcification,vascularity,architectural distortion,duct changes,skin changes,and edema)and BI-RADS category of breast lesions.SE findings were evaluated according to elasticity score(ES)and strain ratio(SR).A total of 4371 patients with solid breast lesions were included in the SE analysis,and three combined methods(BIRADS+ES,BI-RADS+SR,BI-RADS+ES+SR)were used to downgrade BI-RADS category 4a lesions with soft elasticity and to upgrade BI-RADS category 3 lesions with hard elasticity.The predictive values of US features for malignant breast lesions were assessed,and the diagnostic value of BI-RADS incorporated with SE in benign and malignant breast lesions was analyzed by calculating the area under the receiver operating characteristic curve(AUC),sensitivity,specificity,positive predictive value(PPV),negative predictive value(NPV),and accuracy.Results:US features including shape,orientation,margins,echogenic pattern,posterior acoustic features,calcification,vascularity were significantly different in benign and malignant lesions.US features with irregular shape(PPV=63.12%),non-parallel(PPV=72.35%),angular(PPV=78.77%)or spiculated(PPV=91.55%)margins had a high predictive value for malignant lesions.For altered or newly added US features,the PPV was 91.67%,49.10%,and 74.14%for intraductal calcification,internal vascularity,and hard elasticity,respectively.For features associated with breast lesions,architectural distortion(PPV=95.74%),skin changes(PPV=88.33%),and edema(PPV=80.00%)also had a high predictive value for malignant lesions.Compared with the results of BI-RADS(AUC=0.794,sensitivity=99.55%,specificity=59.27%,PPV=57.56%,NPV=99.58%,accuracy=73.64%),BI-RADS+ES had higher AUC(0.830,p<0.0001),specificity(66.70%,p<0.0001),PPV(62.35%,p=0.0004)and accuracy(78.36%,p<0.0001),and similar sensitivity(99.36%,p=0.5078)and NPV(99.47%,p=0.6285);BI-RADS+SR had higher AUC(0.822,p<0.0001),specificity(66.45%,p<0.0001),PPV(61.85%,p=0.0017)and accuracy(77.72%,p<0.0001),but lower sensitivity(98.01%,p<0.0001)and NPV(98.37%,p=0.0004);BI-RADS+ES+SR had higher AUC(0.831,p<0.0001),specificity(66.70%,p<0.0001),PPV(62.36%,p=0.0004)and accuracy(78.38%,p<0.0001),and simliar sensitivity(99.42%,p=0.7266)and NPV(99.52%,p=0.7920).Conclusions:The latest fifth edition of US BI-RADS has good application in clinical practice.Not only the US features of the breast lesion,but also associated features,including architectural distortion,skin changes,edema and elasticity assessment,have become an important part of the fifth edition of BI-RADS US lexicon to distinguish benign from malignant breast lesions.BI-RADS incorporated with SE improved the diagnostic performance in distinguishing benign from malignant breast lesions and could increase specificity and accuracy.Objectives:To establish a breast lesion risk stratification system using B-mode ultrasound(US)images to predict breast malignancy and assess Breast Imaging Reporting and Data System(BI-RADS)categories simultaneously.Methods:This multicenter study prospectively collected a dataset of US images for 5012 patients at thirty-two hospitals from December 2018 to December 2020.A deep learning(DL)model was developed to conduct binary categorization(benign and malignant)and BI-RADS categories(2,3,4a,4b,4c and 5)simultaneously.The training set of 4212 patients and the internal test cohort(ITC)of 416 patients were from thirty hospitals.The remaining two hospitals with 384 patients were used as an external test cohort(ETC).We measured agreement between the results of the DL model and the pathological findings,and the original BI-RADS assessment using kappa values(κ).The diagnostic performance of the DL model was evaluated by the area under the receiver operating characteristic curve(AUC),sensitivity,specificity,positive predictive value(PPV),negative predictive value(NPV),and accuracy.We investigated the assisted value of the DL model by upgrading or downgrading BI-RADS category 3 and 4a lesions.In addition,seven sonographers(one expert,three experienced sonographers,and three inexperienced sonographers)performed Bl-RADS assessments on 324 patients randomly selected from the test sets(comparison set).We compared the performance of the DL model with that of seven sonographers.Results:For DL binary categorization of benign and malignant lesions,the DL model showed substantial agreement(κ=0.759)with the pathology results in the ITC and almost perfect agreement(κ=0.823)in the ETC.For the DL BI-RADS categories,the DL model showed substantial agreement with the sonographers in the ITC and ETC(κ=0.626 andκ=0.669,respectively).In the ETC,the DL model achieved AUCs of 0.980 and 0.945 for the binary categorization and six-way categorizations,respectively.For the DL binary categorization,the sensitivity,specificity,PPV,NPV,and accuracy of the DL model were 89.81%,93.84%,85.09%,95.93%,and 92.71%,respectively;for the DL BI-RADS classification,the sensitivity,specificity,PPV,NPV,and accuracy were 95.37%,82.61%,68.21%,97.85%,and 86.20%,respectively.The DL model was able to improve the diagnostic specificity(76.45%vs.82.25%-93.12%),PPV(62.21%vs.68.39%-84.80%)and accuracy(82.81%vs.86.72%-94.53%)of the sonographers without decreasing the diagnostic sensitivity(99.07%vs.98.15%-99.07%)and NPV(99.53%vs.99.13%99.59%)in ETC.In the comparison set,DL BI-RADS achieved higher average specificity(78.71%vs.66.01%,p<0.0001),PPV(72.08%vs.62.61%,p=0.0014),accuracy(83.33%vs.76.65%,p<0.0001),and similar AUC(0.901 vs.0.911,p=0.3739),sensitivity(90.98%vs.94.26%,p=0.0652),NPV(93.53%vs.95.01%,p=0.3351)than experienced sonographers;and achieved higher average AUC(0.901 vs.0.805,p<0.0001),specificity(78.71%vs.47.52%,p<0.0001),PPV(72.08%vs.51.75%,p<0.0001),accuracy(83.33%vs.64.71%,p<0.0001)and similar sensitivity(90.98%vs.93.17%,p=0.3020),NPV(93.53%vs.92.01%,p=0.4095)than inexperienced sonographers.Conclusions:The DL US-based diagnosis of benign and malignant breast lesions and classification of BI-RADS categories model were established,which achieved excellent diagnostic performance in differentiating benign from malignant breast lesions.The DL model could significantly improve the diagnostic performance of sonographers and increase the accuracy and specificity without decreasing the sensitivity.The DL BI-RADS classification system yielded outcomes similar to experienced sonographers and could mimic sonographers’ decision-making behavior,which indicates the potential applicability of the DL model in clinical diagnosis. |
PDF Full Text Request