Study On The Diagnostic Performance Of Breast Cancer Prediction Model Based On Ultrasound Radiomics

Objective The present study aimed to assess and compare the diagnostic performance of ultrasound-based radiomics models constructed by different machine learning(ML)algorithms for identifying breast cancer.Methods Between January 2017 and April 2019,a total of 828 consecutive patients with pathologically confirmed breast lesions were included and the images of the largest diameter section of the lesion were recorded and stored in DICOM format.Take August31,2018 as the node,set the cases from January 2017 to August 2018 as the training set(n = 526)to build the model,and the cases from September 2018 to April 2019 as the external validation set(n = 302)to test the performance of the model.A sonographer(4years in breast ultrasound)used 3D-Slicer software to trace the contours of the target lesions,and the Pyradiomics software was used to extract ultrasound radiomics features from the target lesions.Two weeks later,60 lesions were randomly selected for re-tracing by the physician and another ultrasound physician(3 years of breast ultrasound)to assess intra-observer and inter-observer reproducibility.Radiomics features were extracted from the gray-scale ultrasound images.After features selection using LASSO regression,five ML models,including k-nearest neighbor,logistics regression,naive bayes,random forest,and support vector machine were constructed.Internal validation was performed using repeated k-fold cross-validation,diagnostic metrics such as sensitivity,specificity,positive predictive value,and negative predictive value were calculated for comparison.Furthermore,model discrimination and calibration were assessed in the external validation.Results 1.A total of 828 cases of breast masses from 828 women were included in the study.359 cases(43.4%)were pathologically diagnosed as benign and 469 cases(56.6%)were malignant.There was no statistically significant difference in the age distribution,the longest diameter of the tumor and the proportion of malignant tumors in the training set and validation set cases(P>0.05);2.19 features from 107 radiomics features were selected as effective features,and 5ML prediction models were established;3.The inter-observer and intra-observer Intraclass correlation coefficient(ICC)range during characterization is 0.754～0.926;4.The internal validation showed statistically difference in all diagnostic metrics among models,the relatively high performance was observed in LR and SVM while relatively low in k NN and NB.5.In external validation,the AUCs of LR,SVM,RF,k NN and NB were 0.890,0.832,0.821,0.746,and 0.703,respectively,indicating good discrimination,the difference between LR and SMV was statistically significant.6.The calibration plot shows that the calibration of LR and SVM was good,while the others were relatively low.Conclusions The ultrasound-based radiomics models established by different ML algorithms showed relatively high diagnostic performance but also some difference,the LR model performed better.Appropriate ML algorithm was conducive to further improving the diagnostic performance of the final model.