Random forest with preoperative core biopsy categories: a novel method for refining ultrasonic Breast Imaging Reporting and Data System evaluation.

Background: Many benign breast lesions are classified as Breast Imaging Reporting and Data System (BI-RADS) category 4, resulting in unneeded biopsies. We thus aimed to build a model based on a core needle biopsy category (CBC) to improve upon BI-RADS classification by analyzing clinical and ultrasonic features.

Methods: A retrospective study was conducted in which female patients with solid breast tumors who underwent ultrasound-guided core needle biopsy (CNB) were enrolled. Participants were randomly allocated to either a training or validation cohort at a 7:3 ratio. We developed CBC prediction models using five machine learning algorithms: support vector machine, random forest (RF), multilayer perceptron (MLP), logistic regression (LR), and k-nearest neighbors (KNNs). The optimal model was selected based on the highest area under the curve (AUC) value and subsequently applied to adjust the BI-RADS categories. The category of BI-RADS was downgraded by one if the CBC prediction was B1 or B2 or upgraded by one if the CBC prediction was B3 or B5. The number and rate of missed or accurate up- or downgrading were calculated.

Results: A total of 1,082 female patients were included comprising 1,185 lesions. The optimal model was RF [AUC =0.943, 95% confidence interval (CI): 0.930-0.956]. In 42 BI-RADS category 3 lesions, 4 (9.5%) cases were upgraded, 3 of which were correct, while 38 (90.5%) cases were downgraded, 37 of which were correct. In 167 BI-RADS category 4A lesions, 149 (89.2%) cases were downgraded, 145 of which were correct, while 18 cases (10.8%) were upgraded, 13 of which were correct.

Conclusions: The predictive model of CBC built by RF can aid in adjusting BI-RADS category 3 and 4A and thus help prevent unnecessary biopsy.