Enhancing Diagnostic Efficiency: A Radiomics Approach for Distinguishing Benign and Malignant Breast Lesions Using BI-RADS Features From Ultrasound Imaging.

Background: Breast cancer is the leading cause of mortality from cancer in women worldwide. Ultrasound is commonly utilized to identify breast cancers but is dependent on operator experience. This study established a radiomics model aimed at enhancing diagnostic efficacy in distinguishing between benign and malignant breast lesions using ultrasound.

Methods: A total of 316 patients were retrospectively included in this study. Two types of feature groups were extracted from ultrasound images: traditional radiomics features and customized features derived from BI-RADS (Breast Imaging Reporting & Data System) classification criteria. The radiomics features were categorized into 3 groups: (A) BI-RADS features, (B) radiomics features, and (C) a combination of both feature groups. Subsequently, SVM (Support Vector Machine), RF (Random Forest) and LR (Logistic Regression) algorithms were utilized to model and analyze based on the 3 feature groups. Finally, the model's performance was evaluated, and the SHAP method was employed to investigate the interpretability of the model.

Results: In Group C, the SVM model demonstrated the best performance on the testing set, achieving an AUC and accuracy of approximately 0.91. The SHAP results revealed that the entropy and variance had the most significant impact on the output of the model (SVM for Group C).

Conclusions: The SVM model constructed using BI-RADS features combined with radiomics feature demonstrated high diagnostic accuracy in distinguishing between benign and malignant breast lesions. This model may assist radiologists in differentiating malignant from benign breast lesions.