| Objective(s):1.To evaluate the BI-RADS classification and benign and malignant classification of breast lesions based on contrast-enhanced ultrasound images,two-dimensional ultrasound images and bimodal ultrasound deep learning model.2.To evaluate the risk of malignant pathology in BI-RADS five classification of breast lesions by ultrasound physicians and deep learning model.3.To evaluate the value of ultrasound imaging in BI-RADS five classification and benign and malignant classification of breast lesions.Methods:The ultrasonographic data and pathological data of patients with breast lesions who underwent puncture biopsy or surgical resection of breast lesions and breast ultrasound examination before operation were collected retrospectively from January2019 to August 2021.A total of 986 breast lesions from 942 patients were included.The images were interpreted by two senior ultrasound doctors with more than 10 years working experience according to the two-dimensional ultrasound images,contrast-enhanced ultrasound images and color Doppler images of the lesions,and with reference to the BI-RADS(breast image reporting and data system)revised by ACR(American Radiological Society)in 2013,and then the model was trained and tested according to the interpretation results.First of all,a deep learning(DL)model is established.Google Net,Moblienet,Resnet101,Xception and Densnet201 are selected to build the deep learning model.According to the pre-experimental results,the network structure with the best diagnosis performance is selected,and then the network structure is used to build the model.Secondly,in the BI-RADS five-classification experiment of breast lesions,the experimental data based on two-dimensional ultrasound images and contrast-enhanced ultrasound images come from 3753 ultrasound images of 986 breast lesions,while the experimental data based on two-dimensional ultrasound combined with contrast-enhanced ultrasound images are derived from fused bimodal images,and the fused data are expanded to 3:4a:4b:4c:5=1100:961:1000:1018:981.In the two-classification experiment of benign and malignant breast lesions: the experimental data based on a single contrast-enhanced ultrasound image came from benign: malignant=2277:1650,and the experimental data based on single two-dimensional ultrasound images were benign: malignant=2150:1603.The accuracy,Precision,F1 score and recall rateof the model were calculated,and the confusion matrix and receiver operating characteristic curve(ROC)were drawn,and the area under the ROC curve(AUC)was calculated to evaluate the predictive performance and clinical practicability of the prediction model.Compared with the pathological results,the results of malignant risk prediction of breast lesions by ultrasound physician reading and optimal depth learning model were compared.Finally,945 cases of breast lesions including two-dimensional ultrasound images,color Doppler images and contrast-enhanced ultrasound images were selected from986 cases of breast lesions.The two-dimensional images,color Doppler images and contrast-enhanced ultrasound images of 945 cases of breast lesions were introduced into the medical quasi-Darwin research platform,and the region of interest(ROI)of the target lesions is delineated layer by layer on the platform,and the benign and malignant lesions and BI-RADS classification are marked.After the delineation is completed,the imaging features of the lesions are selected and extracted,and random number seeds are set to sample according to certain rules.The logical regression model was used for the experiment.The sensitivity,specificity,accuracy,positive predictive value(PPV)and negative predictive value(NPV)of benign and malignant breast lesions and BI-RADS classification were calculated.The ROC curve was drawn and the AUC value was calculated to evaluate the diagnostic performance of the imaging model.Results:1.BI-RADS five-classification experiment results of deep learning model:3753 contrast-enhanced ultrasound images of 986 breast lesions were used to train and test Google Net,Moblienet,Resnet101,Xception and Densnet201 models.Google Net model has the highest diagnostic efficiency among the five deep learning models.The accuracy,accuracy,F1 value,recall rate and AUC of BI-RADS five classification of contrast-enhanced ultrasound images are 94.41%,94.49%,94.42%,94.41% and 96.00%,respectively.3753two-dimensional ultrasound images of breast lesions were used to train and verify the Google Net model.The accuracy,accuracy,F1 value,recall rate and AUC of BI-RADS five classification were 97.34%,97.37%,97.34%,97.34%and 98.00%,respectively.3753 two-dimensional ultrasound images and 3753contrast-enhanced ultrasound images of breast lesions were fused,and after the image data were expanded to 3:4a:4b:4c:5=1100:961:1000:1018:981,the accuracy,accuracy,F1 value,recall rate and AUC of the Google Net model were 90.11%,90.36%,90.08%,90.11% and 94.00%,respectively.Compared with the pathological results,Google Net model classified breast lesions according to BI-RADS five.The malignant proportion of each classification was almost the same as that of ultrasound physicians,except that the malignant proportion of 3 and 4A types was higher than that of ultrasound physicians.2.Experimental results of classification of benign and malignant breast lesions by deep learning model986 cases of contrast-enhanced ultrasound images of breast lesions were augmented and 3927 images were obtained.The augmented images were trained and tested by Google Net model.The accuracy,precision,F1 value,recall rate and AUC of breast lesions were 95.55%,95.62%,95.56%,95.54%and 95.03% respectively.Three thousand seven hundred and fifty-three two-dimensional ultrasound images of 986 cases of breast lesions were used to train and test Google Net model.The accuracy,precision,F1 value,recall rate and AUC of benign and malignant breast lesions were 89.61%,89.63%,89.61%,89.62% and 89.65% respectively.3.Experimental results of BI-RADS five-classification and benignmalignant two-classification of breast lesions by imaging.(1)Benign and malignant two-dimensional ultrasound images,color Doppler images and contrast-enhanced ultrasound images of 945 lesions were each 945.the imaging group model was used to construct the imaging group model.The AUC,ACC,sensitivity and specificity of the model in the training set were 0.902(95% CI:0.879 0.925),0.83,0.832 and 0.892 respectively.The AUC in the test set was 0.908(95% CI:0.873 and 0.943),the ACC was 0.845,the sensitivity was 0.867,and the specificity was 0.83.(2)BI-RADS five classification: There are 934 two-dimensional ultrasound images,934 color Doppler images and 934 contrast-enhanced ultrasound images of 934 lesions,which are used to construct the imaging model.The model has the best classification performance for BI-RADS in the training set,with AUC of 0.915(95% CI: 0.889,0.941),ACC of 0.903,sensitivity of 0.0.562 and.The AUC in the test set is 0.865(95% CI: 0.794,0.937),ACC is 0.909,sensitivity is 0.5,and specificity is 0.958.Conclusion(s):Both deep learning and ultrasound imaging can effectively predict the benign and malignant breast lesions and BI-RADS classification.The main results are as follows:(1)In the BI-RADS classification of breast lesions,the depth learning model based on single contrast-enhanced ultrasound image is more accurate than the model of single two-dimensional ultrasound image and bimodal fusion,and is comparable to the ultrasound physician with 10 years’ working experience.(2)In the classification of benign and malignant breast lesions,the classification performance of the depth learning model based on single contrast-enhanced ultrasound image is better than that of single two-dimensional ultrasound image.(3)The imaging group model based on multimodal ultrasound images has good predictive performance in the classification of benign and malignant masses and has certain clinical application value.(4)In BI-RADS classification,the imaging group model based on multimodal ultrasound images has the best classification performance for 5and 4C lesions. |
PDF Full Text Request