Value Of Radiomics Based On Conventional Ultrasound Imaging In Differentiating Malignant From Benign Breast Tumors

Objective Breast cancer has the highest incidence rate among female cancers in the world,and its mortality rate ranks second.Early and accurate diagnosis of breast cancer can help improve the patient’s prognosis.At present,the most widely used diagnostic criteria of breast conventional ultrasound images is the breast imaging report and data system,which mainly distinguishes benign and malignant tumors based on morphological information such as tumor morphology,margins,and internal echo.The diagnostic accuracy is high,but subjectivity.Radiomics can transform medical images into high-dimensional,available quantitative radiomics features through high-throughput data feature extraction algorithms,and use various algorithms to conduct depth mining and analysis on the extracted radiomics features.In this study,a predictive model of radiomics based on gray-scale ultrasound of breast tumors was constructed to investigate the value of radiomics based on conventional ultrasound imaging in differentiating malignant from benign breast tumors,and the diagnostic efficiency of this model was compared with sonographers with different working experiences.MethodsA analysis was performed for the breast mass ultrasound imaging of 362 patients in our hospital who had been confirmed pathologically from December 2018 to December2019.Tumor lesions were segmented and a large number of radiomics features from ROIs were extracted.A total of 396 radiomics features were extracted from conventional ultrasound images.Then the data set was randomly divided into imaging radiomics training cohort(n=241)and validation cohort(n=121).The minimum Redundancy Maximum Relevance(m RMR)and the least absolute shrinkage and selection operator(LASSO)was applied to select the most important features to build radiomics model.Receiver operating characteristic curve(ROC)and calibration curve were utilized to evaluate the performance of the model.The same methods were used to test the predictive performance of the radiomics model in an independent validation cohort.In addition,this model was compared with the performances of ultrasound doctors with different working experience.ResultsA total of 15 optimal features were selected to form radiomics model in the end.The area under the ROC curve,accuracy,sensitivity,specificity for the training cohort were0.84,0.760,0.839,0.713,and the area under the ROC curve(AUC),accuracy,sensitivity,specificity for the validation cohort were 0.84,0.714,0.821,0.652.The calibration curve showed that there was no significant deviation between the prediction result and the model prediction(P>0.05).Performance of radiomics model reached the level of doctor with 10-12 years experience(P>0.05),was lower than the performance of doctor with 25-30 years experience(P<0.001),but was better than the performance of resident with 4-5 years experience(P<0.001).ConclusionRadiomics model based on conventional ultrasound performs well in the differentiation of malignant from benign lesions.