Applied Research Of AI-based Ultrasonic Imaging In Breast Masses Detection And Benign As Well As Malignant Differential Diagnosis

Objective:To explore the application value of AI technology in the intelligent detection and benign as well as malignant diagnosis of breast masses in ultrasonic imaging,and to verify whether AI technology can help improve the level of breast masses differential diagnosis of benign and malignant by primary and intermediate physician.Methods:Retrospectively collected 401 patients with breast masses which were confirmed by biopsy,surgical pathology,etc.as the main causes of seeking medical advice at Zhongshan Hospital affiliated to Dalian University from June 2017 to April 2019.Among them,204 cases were benign?accounting for 50.9%?,197 cases were malignant?accounting for 49.1%?,ranging in age from 18 to 89 years old,with an average age of48.8±13.9 years old.After the original ultrasonic images?including 582 single images and jigsaw images in total?were collected,they were preprocessed by AI researcher,then a chief physician and a resident physician jointly labeled all breast masses in the 697 images which were preprocessed and obtained before,and randomly selected 348 images as part of the training set of this study?other ultrasonic images in the training set were from other hospitals?,and the remaining 349 images were used as the test set.The AI model which based on deep learning?DL?method adopted in this study is an improved RFBNet model of convolutional neural network?CNN?.Next,a resident physician?the same resident physician who labeled the masses?and an attending physician diagnosed the 349 test set images without the aid of AI prediction results,and at the same time,AI researcher used3568 ultrasonic images provided by other hospitals and 348 images randomly selected by our hospital to jointly train the AI model used in this research,and then used the model to detect the 349 test set images and obtain the results predicted by the model.Finally,the resident physician diagnosed the 349 test set images again with the aid of AI prediction results?according to the independent diagnosis results of the attending physician,the multiple parameter indexes were higher than the AI prediction results,so the attending physician in this study failed to diagnose the 349 test set images again with the aid of AI prediction results?.In this study,SPSS 21.0 statistical software was used for data analysis and processing.The differences between groups of parameter indexes were compared by McNemar test?paired chi-square test?.ROC curve was used to analyze the efficiency of AI prediction,independent diagnosis of the resident physician,AI-assisted diagnosis of the resident physician,and independent diagnosis of the attending physician.P<0.05 as the difference was statistically significant.Results:?1?In this study,AI predicts that the true positive rate?TPR?is 82.4%,the positive predictive value?PPV?is 78.4%,the false positive rate?FPR?is 23.1%,the false discovery rate?FDR?is 21.6%,the false negative rate?FNR?is 17.6%,the negative predictive value?NPV?is 81.1%,the true negative rate?TNR?is 76.9%,the negative omission rate?FOR?is 18.9%,the average accuracy?AP?is 79.7%,the F_0.5-score.5-score is 79.1%,the F_1-score-score is 80.3%,the F_2-score-score is 81.6%,the accuracy rate is 79.7%,while the accuracy rate of AI for detecting the location of breast masses is 100%.?2?The resident physician diagnoses independently that the TPR is 88.6%,the PPV is 74.6%,the FPR is 30.6%,the FDR is 25.4%,the FNR is 11.4%,the NPV is 85.7%,the TNR is 69.4%,the FOR is 14.3%,the AP is 80.2%,the F_0.5-score.5-score is 77.1%,the F_1-score-score is 81.0%,the F_2-score-score is 85.4%,and the accuracy rate is 79.1%.?3?With the help of AI,the resident physician diagnoses that the TPR is 83.0%,the PPV is 77.2%,the FPR is 24.9%,the FDR is 22.8%,the FNR is 17.0%,the NPV is 81.3%,the TNR is 75.1%,the FOR is 18.7%,the AP is 79.2%,the F_0.5-score.5-score is78.3%,the F_1-score-score is 80.0%,the F_2-score-score is 81.7%,and the accuracy rate is 79.1%.?4?The attending physician diagnoses independently that the TPR is 81.3%,the PPV is 85.1%,the FPR is 14.5%,the FDR is 14.9%,the FNR is 18.7%,the NPV is 81.8%,the TNR is 85.5%,the FOR is 18.2%,the AP is 83.4%,the F_0.5-score.5-score is 84.3%,the F_1-score-score is 83.1%,the F_2-score-score is82.0%,and the accuracy rate is 83.4%.?5?There were no statistically significant differences for TPR and TNR between AI-assisted diagnosis of the resident physician and independent diagnosis of the resident physician?P>0.05?.The differences of TPR and TNR between independent diagnosis of the attending physician and independent diagnosis of the resident physician were statistically significant?P<0.05?,and the difference of accuracy rate was not statistically significant?P>0.05?.?6?The area under the curves?AUC?of AI prediction,independent diagnosis of the resident physician,AI-assisted diagnosis of the resident physician,and independent diagnosis of the attending physician were 0.796,0.790,0.790 and 0.868 respectively.Conclusion:?1?The AI model which based on DL method used in breast masses detection as well as classification diagnosis of benign and malignant in this study is an improved RFBNet model of CNN,the results show that the accuracy rate of detection and classification as well as other indicators of the AI model are ideal,which can provide some auxiliary and reference value for clinical ultrasonic diagnosis of breast masses.?2?According to the current AI model adopted in this study,its diagnosis efficiency can not further improve the level of breast masses differential diagnosis of benign and malignant by primary and intermediate physician.