Predicting female breast cancer by artificial intelligence: Combining clinical information and BI-RADS ultrasound descriptors

Abstract

This study aimed to use Artificial Intelligence (AI) Deep Learning (DL) techniques to predict female breast cancer detected by ultrasound based on clinical data and Breast Imaging Reporting and Data System (BI-RADS) Ultrasound (US) descriptors. We retrospectively gathered data on clinical information and BI-RADS US descriptors of breast lesions from 1051 female patients, forming a comprehensive dataset. Two datasets (A and B) were derived by selecting different variables. A BI-RADS DL-based Network (BD-Net) was developed and trained on Dataset A and B, and its performance was evaluated on an external test set. Radiologists also classified Dataset B and the external test set using BI-RADS US. Performance in predicting the probability of malignancy was evaluated by calculating the Area Under Curve (AUC), accuracy, sensitivity, and specificity. BD-Net achieved an accuracy of 92.5% (95%CI, 90.5–94.2) in predicting breast cancer with a sensitivity of 93.0% (95%CI, 90.3–95.4), a specificity of 92.1% (95%CI, 89.7–94.6), and an AUC of 0.97 (95%CI, 0.96–0.98) on the training data set of dataset A. On the external dataset, the BD-Net showed a sensitivity of 93.8% (95%CI, 87.5–98.8), a specificity of 91.0% (95%CI, 85.0–96.0), and an AUC of 0.92 (95%CI, 0.88–0.97) for predicting breast cancer. The radiologists predicted breast cancer on Dataset B and the external test set with AUC values between 0.75 (95%CI, 0.75–0.79) and 0.82 (95%CI, 0.77–0.87). These results indicate that the BD-Net is effective for predicting ultrasound-detected female breast cancer.