Establishment of prediction model for breast lesion using automated breast ultrasound system

Automated breast ultrasound systems (ABUS) have been increasingly used for breast lesion detection; however, standardized and interpretable prediction models based on ABUS lexicon features for differentiating benign from malignant lesions remain limited. Therefore, this study aimed to establish a prediction model that differentiates malignant from benign breast lesions using an ABUS.

This prospective study enrolled patients with single breast lesions identified by handheld ultrasound (HHUS) in Peking University Cancer Hospital between June 2010 and December 2012. The ABUS was performed and the ultrasonic features were described based on the fifth edition of the Breast Imaging Reporting and Data System (BI-RADS) ultrasound lexicon and related literature of lesions and a pathology examination was performed to confirm pathological status. The model performance was evaluated using the area under the curve (AUC).

There were 2,090 patients with single breast lesions included and 630 cases (30.1%) were selected for pre-test. Patients were then randomly split 1:1 into a training set (n=315) and a validation set (n=315). Univariate analysis revealed that lesion form, shape, orientation, margin, boundary, and posterior acoustic feature were significantly different between the benign lesion and malignant lesion group (all P<0.05). The lesion form, boundary, and margin were further selected for multivariate model development. The equation was Y=1.604 × boundary + 1.045 × lesion form - 5.436 × margin (A) - 2.166 × margin (B), the AUC reached 0.882 [95% confidence interval (CI): 0.844-0.919] in training set and 0.866 (95% CI: 0.824-0.928) in validation set.

Lesion form, boundary, and margin information might be associated with benign and malignant lesions on ABUS. The malignant lesion prediction model based on lesion form, boundary, and margin showed great diagnostic performance in both training and validation cohorts.