Machine learning based on automated breast volume scanner radiomics for differential diagnosis of benign and malignant BI-RADS 4 lesions / 中华超声影像学杂志

ABSTRACT

Objective:To evaluate the performance of machine learning (ML) based on automated breast volume scanner (ABVS) radiomics in distinguishing benign and malignant BI-RADS 4 lesions.Methods:Between May to December 2020, patients with BI-RADS 4 lesions from the Affiliated Hospital of Southwest Medical University (Center 1) and Guangdong Provincial Hospital of Traditional Chinese Medicine (Center 2) were prospectively collected and divided into training cohort (Center 1) and external validation cohort (Center 2). The radiomics features of BI-RADS 4 lesions were extracted from the axial, sagittal and coronal ABVS images by MaZda software. In the training cohort, 7 feature selection methods and thirteen ML algorithms were combined in pairs to construct different ML models, and the 6 models with the best performance were verified in the external validation cohort to determine the final ML model. Finally, the diagnostic performance and confidence (5-point scale) of radiologists (R1, R2 and R3, with 3, 6 and 10 years of experience, respectively) with or without the model were evaluated.Results:①A total of 251 BI-RADS 4 lesions were enrolled, including 178 lesions (91 benign, 87 malignant) in the training cohort and 73 lesions (44 benign, 29 malignant) in the external validation cases. ②In the training cohort, the 6 ML models (DNN-RFE, AdaBoost-RFE, LR-RFE, LDA-RFE, Bagging-RFE and SVM-RFE) had the best diagnostic performance, and their AUCs were 0.972, 0.969, 0.968, 0.968, 0.967 and 0.962, respectively. ③In the external validation cohort, the DNN-RFE still had the best performance with the AUC, accuracy, sensitivity, specificity, PPV and NPV were 0.886, 0.836, 0.934, 0.776, 86.8% and 82.5%, respectively. ④Before model assistance, R1 had the worst diagnostic performance with the accuracy, sensitivity, specificity, PPV and NPV were 0.680, 0.750, 0.640, 57% and 81%, respectively. After model assistance, the diagnostic performance of R1 was significantly improved ( P=0.039), and its diagnostic sensitivity, specificity, accuracy, PPV and NPV improved to 0.730, 0.810, 0.930, 68% and 94%; while the improvement of R2 and R3 were not significantly ( P=0.811, 0.752). Meanwhile, the overall diagnostic confidence of the 3 radiologists increased from 2.8 to 3.3 ( P<0.001). Conclusions:The performance of ML based on ABVS radiomics in distinguishing between benign and malignant BI-RADS 4 lesions is good, which may improve the diagnostic performance of inexperienced radiologists and enhance diagnostic confidence.