| Objective:Based on the postoperative pathological results of breast cancer as the gold standard,on the subjective characteristics of ultrasound images and the data processed by different deep learning models and the open platform of Baidu Brain(EasyDL),to explore the value of prediction of breast cancer in BI-RADS 4 breast lesions using different deep learning models,EasyDL,and Nomogram to establish breast-ultrasound-related models.Methods:1 Retrospectively collected 17231 ultrasound images of female patients undergoing breast cancer resection from January 2019 to December 2019 in the Third Affiliated Hospital of Kunming Medical University(2578 cases).After rigorous screening,age,menarche age,first-class family history,preoperative ultrasound images,descriptive reports,and postoperative pathological data were obtained.After inclusion criteria,exclusion criteria and image preprocessing,717 ultrasound images of breast tumors were included in this study,a total of 260 patients.2 By two doctors who have over-10-year experience in breast ultrasound diagnosis evaluating independently through the characteristics of breast ultrasound image classification of benign and malignant,when the opinion is not unified by a chief physician final decision.According to the 2019 edition of the guidelines and guidelines for the diagnosis and treatment of breast cancer of the Chinese anticancer association,the evaluation included in this study included:shape of the mass,aspect ratio,boundary,margin,echo pattern,posterior echo of the lesion,surrounding tissue and calcification.30 patients were randomly selected by computer,evaluated by two doctors,and the consistency of the two doctors was evaluated by Kappa test.3 Establish deep learning model.The transfer learning model ResNet18,ResNet34,ResNet50 and ResNet101 were selected as the pre-training model of this study,which was divided into the following steps according to the steps:data preparation,data expansion and enhancement,data loading,transfer learning,training,verification,testing.Baidu brain(EasyDL)model through platform training,verification,testing to complete the construction of the exclusive model.4 The shape,aspect ratio,boundary,edge,echo pattern,posterior echo,surrounding tissue and calcification of all cases were included in multivariate Logistic regression analysis,and the statistically different breast ultrasound features were included in R language and a single physician subjective evaluation line map prediction model was established.5 The age,initial age,and level I family history of all cases were included in multiple Logistic regression analysis for screening,and the statistical clinical risk factors selected for clinical risk factors were built combined with the above statistical characteristics of breast ultrasound.6 The following part is referred to as the machine clinical ultrasound prediction model.The deep learning model derived the real results,prediction results and prediction probability of each ultrasound image,predicted the outcome of 30 images combined with the best deep learning models,and jointly selected the optimal physician subjective evaluation model to obtain a joint deep learning-physician subjective evaluation model.7 ROC curves were drawn for all 8 models to obtain AUC,accuracy,specificity,sensitivity,specificity,negative predictive value,positive predictive value,recall rate,F1 value and Kappa value.Finally,we will compare all the models.Results:1 General data:260 patients(717 ultrasound images of breast)classified as BI-RADS 4 with definite pathological findings due to breast lumps,Of these,154 patients were benign,106 patients with malignancy.There were 113 cases,28 cases and 13 cases of benign lesion of BI-RADS 4A,4B and 4C,There were 36 cases,41 cases and 29 cases of malignant lesion of BI-RADS 4A,4B and 4C.Among them,119 cases of BI-RADS 4A,47 cases of 4B and 29 cases of 4C lesions were correctly judged by physicians,while 30 cases,5 cases and 13 cases were wrongly judged.The pathological types of benign patients were 83 cases of fibroadenoma,40 cases of mastopathy(including sclerosing adenosis),18 cases of intraductal papilloma,10 cases of acute and chronic inflammation with granulomatous lobular mastitwas,2 cases of benign lobular tumor,1 case of fibromuscular myofibroblast-derived tumor,The pathological types of malignant patients were invasive ductal carcinoma 79 cases,intraductal carcinoma in situ 24 cases,mucinous carcinoma 3 cases.Average age(45.14±9.80)years,Among them,the average age of patients with benign pathology(41.96±8.66)years,Average age of malignant patients(49.76±9.56)years.Age difference between benign and malignant groups was statistically significant(P<0.05).Average age of menarche(13.64±1.72)years,Among them pathological result was benign patient menarche average age(13.68±1.50)years old,Average age of menarche in malignant patients(13.58±2.00)years.Between benign and malignant groups,there was no significant difference in menarche age(P>0.05).There were 236 patients(90.8%)with no family history of malignant tumor,There were 23 patients with a family history of primary malignant tumors(8.8%).More than 1 case of family history of primary malignant tumor(0.4%).Among them,149 patients with no family history of malignant tumor(96.8%),5 patients with a family history of primary malignancy(3.2%),More than 1 patients with family history of primary malignant tumor 0 cases(0%).87 patients with no family history of malignancy(82.1%),18 patients(17.0%)with a family history of primary malignant tumors,More than 1 case(0.9%)with family history of primary malignant tumor,The mother,a sister and a sister all had breast cancer.A family history was statistically significant between benign and malignant groups(P<0.05).Among the clinical risk factors,the difference between benign and malignant groups was statistically significant(P<0.05).2 Comparison of breast ultrasound features:the difference between benign and malignant groups was statistically significant:shape,aspect ratio,boundary,edge,internal echo,posterior echo of lesion and calcification(P<0.05).3 The consistency test of the subjective evaluation of the two doctors:the two doctors conducted a consistency test on the subjective evaluation of the ultrasound characteristics of the breast masses in 30 patients randomly selected.The results showed that only in the breast mass "internal echo" was not consistent,other indicators were consistent.The ultrasonic features with good consistency were:echo behind the lesion.Ultrasonic features with good consistency were shape,boundary,edge and calcification.Ultrasonic features with moderate consistency were peripheral tissue invasion and aspect ratio.4 To establish a single physician subjective assessment model and clinical risk factors-doctor subjective assessment model:the characteristics of the Logistic multivariate regression model were screened twice,and the statistically significant ultrasonic characteristics were:aspect ratio,boundary,echo behind the lesion and calcification(P<0.05).The statistically significant clinical risk factors were age,first class family history.Through the construction of Logistic multi-factor regression model,and then using R language to establish a single physician subjective evaluation model and clinical risk factors-doctor subjective evaluation model.From the ROC curves of the two models,A single physician subjective assessment model AUC 0.809,The sensitivity was 81.25%,The specificity was 92.93%,The prediction accuracy of the model was 82.31%.Clinical risk factors-the AUC of the physician’s subjective assessment model was 0.845,The sensitivity was 81.31%,The specificity was 87.58%,The prediction accuracy of the model was 85.00,Both models have good predictive performance,But clinical risk factors-the subjective assessment model of physicians was more convincing.From the calibration curves of the two models,Both models have good predictive power,But there was a better coincidence between the predictive value and the observed value of the single physician’s subjective evaluation model,The accuracy of the prediction model was higher.The curve(probility)of the two models was higher than that of the extremum curve(All and None),which indicates that both models have clinical application value.However,the curve of clinical risk factor-doctor subjective evaluation model was higher than that of single physician subjective evaluation model,which reflects higher clinical application value.5 establish a combination model of deep learning,clinical risk factors and physician’s subjective assessment:combine ResNet50 and clinical risk factors and physician’s subjective assessment model to establish a combined model of deep learning,clinical risk factors and physician’s subjective assessment.6 Compare all the models.The result of deep learning model ResNetl8,ResNet34,ResNet50,ResNet101,single physician subjective assessment model,clinical risk factor-doctor subjective assessment model,deep learning-clinical risk factor-doctor subjective assessment model are as follow:AUC were 0.748,0.851,0.856 and 0.767,0.809,0.849 and 0.90,respectively.Accuracy(%)were 76.85,86.19,86.89,78.24,82.31,85.00,90.00.Precision(%)were:88.78,92.44,94.39,87.56,73.59,82.08,86.67.Sensitivity(%):75.21,84.79,84.50,77.37,81.25,81.31,92.86.Specificities(%)were:80.26,88.52,91.12,79.84,85.00,82.93,87.58,87.50.Positive predictive values(%)were:88.78,92.44,94.39,87.56,73.59,82.08,86.67.Negative predictive values(%)were:60.91,77.85,76.87,65.80,88.31,87.01,93.33.Recall rates(%)were:75.21,84.79,84.50,77.37,81.80,81.25,81.31,92.86.F1 values(%)were:81.43,88.45,89.17,82.15,81.60,77.23,81.69,89.66.Kappa values were:0.512,0.714,0.727,0,546,0.628,0.690,0.800.The accuracy(%)of EasyDL was 82.30,the accuracy rate(%)was 81,50,sensitivity(%)was 78.00,specificity(%)was 85.00,recall rate(%)was 81.80,F1 value(%)was 81.60.7 The highest AUC(AUC=0.900)of the combined model of deep learning-clinical risk factors-mechanics subjective assessment showed that the overall prediction ability of the model was the best,followed by the better prediction abilityResNet50(AUC=0.856).The Combined Model of Deep Learning-Clinical risk Factors-Physicians’ Subjective Assessment got high sensitivity(92.86%),specificity(87.50%)and lowest missed diagnosis rate.ResNet50 rate The highest specificity(91.12%),Medium sensitivity(84.50%),The rate of misdiagnosis of the model was the lowest,and the accuracy rate of the combined model of deep learning-clinical risk factors-mechanics subjective assessment was the highest(90.00)in all models%The highest accuracy in the line chart model was the clinical risk factor-the physician’s subjective assessment model(85.00%),and the highest accuracy in the deep learning modelResNet50(86.89%).ResNet50 had the highest positive predictive value(94.39%),and the model had the highest efficacy in predicting malignant cases correctly,and the negative predictive value(93.33%)of the combined model of deep learning-clinical risk factors-doctor subjective assessment was the highest.ResNet50 accuracy rate was the highest(94.39%),and the model was the best in predicting benign and malignant lesions correctly among all the predicted results.The recall rate of the combined model of deep learning-clinical risk factors-doctor subjective assessment was the highest(92.86%),and the model was the best in predicting benign and malignant lesions correctly.The F1 value and Kappa value of the combined model of deep learning-clinical risk factors-doctor subjective assessment were the highest(89.66 and 0.800),and the accuracy of the model was the best.Baidu Brain(EasyDL)was no worst indicators for each set of data,which basically meet the requirements of the deep learning model.Conclusion:1 Compared with the subjective assessment model and clinical risk factor-subjective assessment model,the deep learning model based on ultrasound was valuable in predicting breast cancer in BI-RADS 4 kinds of breast lesions,which can improve the sensitivity of breast cancer prediction.2 Compared with Logistic traditional regression analysis equation,the 2-Nomogram graph has more readability and better prediction efficiency.3 Although the amount of data for deep learning should be more in theory,but in the case of small amount of data also obtained a higher accuracy,if continue to improve the amount of data,then the model prediction effectiveness will be better reflected,prove that the ResNet deep-learning model was suitable for ultrasonic image classification applications.4 The 400 degree brain(EasyDL)platform has very perfect function for image classification at present.Under the environment of less data and free configuration of the platform in this study,the accuracy rate comparable to that of the ResNet can be obtained.It proves that EasyDL can provide a reliable AI platform for the basic area,primary learners or zero-base users.5 Although AI become the focus in the medical area,but we still need to put the human being in the indispensable place and dialectically look at the relationship between AI and clinical practice,which can help us grasp the future better and better. |
PDF Full Text Request