| Part one Research on the value of ACR US-BI-RADS term features in BI-RADS 3-5 classification accuracy evaluation based on decision treeObjective: Using the 2013 edition of ultrasound BI-RADS dictionary to describe the standardized term features of ultrasound images of breast lesions,and the final BIRADS evaluation classification was given.A text data was generated,and single factor and binary logistic regression analysis were used to elect the term features which were different between benignancy and malignancy,and different scores was put in the term features.The sum of all term feature scores was used for 10-fold cross validation after machine learning of decision tree,and the total scores are divided into five categories of BI-RADS according to different ranges,including 3,4A,4B,4C,5,which provides a quantitative index for BI-RADS 3 to 5 evaluation classification.Methods: The study breast ultrasound images and text data were collected from the workstation database of US reports from the ultrasound medicine department of West China hospital of Sichuan University from March 2017 to December 2019,and all cases had one-to-one corresponding pathological results.Breast US examinations were analyzed by three senior radiologists.The ultrasound images of breast lesions were described according to the 2013 BI-RADS edition,and a proper BI-RADS category was recorded.Univariate regression analysis was used to elect the term features which were significant differences between benign and malignant lesions,and binary logistic regression was performed with all the selected term features.Finally,each lesion would be given an assignment which was the sum of all the term features scores.BI-RADS category was classified with the method of Decision tree 10-fold cross validation.The positive predictive value of lesions with different scores was estimated.ROC with calculation of AUC was used to evaluate diagnostic ability,the cut-off value was also calculated.Result: A total of 2500 cases of breast lesions were finally included.500 cases were classified in category 3,500 cases in category 4A,500 cases in category 4B,500 cases in category 4C,500 cases in category 5.1377 lesions were benign,and the remaining 1123 lesions were malignant.All the patients were female,age ranged from 14 to 89 years old with a mean of 45.7±11.5.Patients with benign lesions ranged in age from 14 to 89 years old with a mean of 41.3±10.3,598 cases were less than 40 years old,779 cases were 40 years old and above.Patients with malignant lesions ranged in age from 24 to 89 years old,125 cases were less than 40 years old,998 cases were 40 years old and above.The age of patients with lesions was significantly higher than that of patients with benign lesions(P<0.001).The PPV of BI-RADS category 3 was 1.6%,4A was 5.8%,4B was 38.6%,4C was 81.6%,and 5 was 97.2%.Term features of irregular shape,nonparallel,indistinct margin,angled margin,micro-lobulated margin,speculated margin,post shadowing,calcification(micro-calcification,macrocalcification),structure distortion,edema,marked intro-nodular blood flow,abnormal axillary lymph node was significant different between benignancy and malignancy.According to the result of binary logistic regression,these above lexicons were assigned in turns as follows: 2 point,1point,1 point,1point,1 point,1 point,2point of micro-calcification,-1 point of macro-calcification,1 point,1point,1 point,1 point.Decision tree 10-fold cross validation was used to verify the scores of each category.The score of each category of 3,4A,4B,4C and 5 was no more than 0,1 to 2 point,3 to 5 point,6 to 7 point and ≥ 8,and the PPV was 0.43%,5.44%,34.92%,85.56% and 98.38% in turns,the AUC was 0.951.The cut-off value was ≥ 5,and the sensitivity,specificity and accuracy were 87.6%,89.5% and 89.9%,respectively.Conclusion: BI-RADS has an important value for the standardized description of breast nodules of ultrasonic images,and the assessment was in accordance with the malignant risk.When the lexicons were assigned,the larger of the sum of different characteristic scores,the higher the malignant risks were.The score of each category of 3,4A,4B,4C and 5 was no more than 0,1 to 2 point,3 to 5 point,6 to 7 point and ≥ 8,the malignant risk of each category was reasonable.Part two Application of convolutional neural network(CNN)in feature extraction and prediction of BI-RADS 3-5 Classification of breast ultrasound imagesObjective: Using breast ultrasound images,based on the training set,different models of convolutional neural network algorithm are established to classify BI-RADS 3-5.The test set data are used to verify the prediction performance of different CNN methods.Methods: According to the results of part one,the BI-RADS classification results of different breast ultrasound images by ultrasound doctors are taken as the gold standard,and each breast ultrasound image is labeled with BI-RADS classification,that is,each breast ultrasound image corresponds to an classification result.In this part,the lesion ultrasound images were segmented based on deep learning,and the irrelevant information was removed.The pathologic areas were outlined with dimensioning software.All the cases were divided into two groups,one was training group,the other was test group.Training group was analyzed with U-net split network,VGG 16 network model,VGG 19 network model,Resnet 50 network model and Perception V3 network model.5-fold cross validation was used to establish prediction models of BI-RADS assessment category.Test group was used to verify the efficiency of these models.The sensitivity,specificity,accuracy,positive predictive value,negative predictive value,misdiagnosis rate.When P<0.05,the difference was statistically significant.Result: U-Net segmentation network established the prediction model through the training set,and carried out 5-fold cross validation on extraction for ROI of breast lesions in the test set.The average dice results of ROC segmentation were 84.1%,86.1%,86.5%,85.5% and 84.6%,respectively.The AUC of BI-RADS category 3,4A,4B,4C,5 was 0.936,0.809,0.852,0.862,0.976 in turns when using VGG 16 baseline model.They were 0.938,0.812,0.857,0.900,0.981 in turns when using VGG 19 baseline model.When Res Net 50 baseline model was used,they were 0.960,0.847,0.859,0.847,0.978 in turns.They were 0.962,0.866,0.906,0.874,0.981 in turns when using Inception V3 baseline model.The difference among Inception V3 baseline model and VGG 16 and VGG19 baseline model were statistical significance(P<0.05).The AUC of VGG16 baseline model combined ROI model to predict BIRADS category 3,4A,4B,4C,5 was 0.962、0.866、0.906、0.874、0.981.The AUC of VGG19 baseline model combined ROI model to predict BI-RADS category 3,4A,4B,4C,5 was 0.949,0.819,0.898,0.877,0.982.The AUC of Res Net50 baseline model combined ROI model to predict BI-RADS category 3,4A,4B,4C,5 was 0.961,0.842,0.910,0.871,0.978.The AUC of Res Net50 baseline model combined ROI model to predict BI-RADS category 3,4A,4B,4C,5 was 0.961,0.842,0.910,0.871,0.978.The AUC of Inception V3 baseline model combined ROI model to predict BIRADS category 3,4A,4B,4C,5 was 0.968,0.860,0.929,0.873,0.977.The difference between Inception V3+ROC model and VGG19+ROC model,Resnet50+ROC model was statistical significance(P<0.05).Inception V3 baseline combined ROI model was the best based on the Youden index.The sensitivity,specificity,accuracy,positive predictive value,negative predictive value,misdiagnosis rate,missed diagnosis rate of BI-RADS category 3 were 92.2%,91%,91%,85.2%,94.8%,9%,8.9%.For category 4A,They were 87.4%,74.1%,75.7%,39%、96.6%,25.9%,15.3% in turns.For category 4B,they were 91.8%,83.2%,83.4%,27.7%,98.8%,16.8%,14.1%.For category 4C,they were 98.2%,74.8%,75.9%,29.6%,97.8%,25.2%,14.6%.For category 5,they were 98.2%,89.9%,92.1%,81.8%,98.4%,10.1%,3%.Conclusion: U-Net segmentation prediction model can better extract the ROI of breast ultrasound images,and the accuracy is more than 84%.For BI-RADS 3-5 classification breast ultrasound images,whether it is the baseline model or the baseline + ROI model,the classification efficiency of perception V3 is the best,and its accuracy of predicting BI-RADS 3,4A,4B,4C,5 categories can reach 91%,75.7%,83.4%,75.9%,92.1%,respectively.When predicting BI-RADS 4 category lesions,the accuracy rate was significantly lower than that of 3 and 5 category lesions,and the misdiagnosis rate and missed diagnosis rate were also significantly increased.The reason of the malignant rate of 4 types of lesions varies widely with 3-94%,which was why it was difficulty of accurate classification of 4 sub-category classification.Part three BI-RADS term features combined with convolution neural network(CNN)of breast ultrasound image for BI-RADS 3-5 classificationObjective: Application of deep learning model in BI-RADS of breast ultrasound image 3-5 classification shows good classification effect for 3 and 5 categories.However,the efficiency of BI-RADS category 4A,4B,4C classification needs to be improved.The term feature description of breast lesions by ultrasound doctors is coded,and the deep learning model is added together with the corresponding breast ultrasound images,in order to improve the accuracy of different classification.Methods: The BI-RADS term features in breast ultrasound images were encoded by computer,and combined with U-Net,VGG16,VGG19,Res Net 50 and Inception V3 in the second part to form a baseline + ROI + term feature model for deep learning.As in the second part,2250 cases in the training set were trained,and 250 cases in the test set were used for classification prediction.ROC curve was drawn to obtain AUC,and sensitivity,specificity,accuracy,positive predictive value,negative predictive value,misdiagnosis rate and missed diagnosis rate of different categories were obtained.P < 0.05 means there is statistical difference.Results: The AUC of BI-RADS category 3,4A,4B,4C,5 was 0.999,0.995,0.998,0.968,0.989 in turns when using VGG 16 baseline combined ROI and text data model.They were 1.000,0.996,0.999,0.971,0.989 in turns when using VGG 19 baseline combined ROI and text data model.When Res Net 50 baseline combined ROI and text data model was used,they were 0.996,0.970,0.990,0.960,0.985 in turns.They were 0.999,0.975,0.981,0.940,0.983 in turns when using Inception V3 baseline combined ROI and text data model.The difference among Inception V3 baseline combined ROI and text data model and VGG 16 and VGG19 baseline combined ROI and text data model were statistical significance(P<0.05).VGG19 baseline combined ROI and text data model was the best based on the Youden index.The sensitivity,specificity,accuracy,positive predictive value,negative predictive value,misdiagnosis rate,missed diagnosis rate of BI-RADS category 3 were 99.6%,99.4%,99%,98.9%,99.1%,0.6%,1.6%.For category 4A,They were 99.5%,97.9%,97.8%,89.7%、99.4%,2.1%,3.2% in turns.For category 4B,they were 100%,99.2%,98.9%,92.2%,99.6%,0.8%,5.9%.For category 4C,they were 97.7%,88.8%,89.4%,51.6%,99.2%,11.2%,6.2%.For category 5,they were 99.2%,93.2%,94.7%,87.0%,99.0%,6.8%,2%.Conclusion: VGG19 baseline combined ROI and text data model was the best in BI-RADS assessment category compared with the other three models.The accuracy of prediction of BI-RADS category 3,4A,4B,4C,5 with VGG19 baseline combined ROI and text data model was 99%,97.8%,98.9%,89.4%,94.7%,respectively.The accuracy of BI-RADS 4 sub-type prediction was significantly improved,besides,the misdiagnosis rate and missed diagnosis rate were also significantly decreased.Deep learning is not only used for image data,but also for text data.The combination of image and text data could effectively improve the classification efficiency,which is worthy of further research in the future. |
PDF Full Text Request