Machine-learning models for differentiating benign and malignant breast masses: Integrating automated breast volume scanning intra-tumoral, peri-tumoral features, and clinical information

Differentiating between benign and malignant breast masses is critical for clinical decision-making. Automated breast volume scanning (ABVS) provides high-resolution three-dimensional imaging, addressing the limitations of conventional ultrasound. However, the impact of peritumoral region size on predictive performance has not been systematically studied. This study aims to optimize diagnostic performance by integrating radiomics features and clinical data using multiple machine-learning models.

This retrospective study included ABVS images and clinical data from 250 patients with breast masses. Radiomics features were extracted from both intratumoral and peritumoral regions (5, 10, and 20 mm). These features, combined with clinical data, were used to develop models based on four algorithms: Support vector machine, random forest, extreme gradient boosting, and light gradient boosting machine (LGBM). Model performance was evaluated using area under the receiver operating characteristic curve (AUC), calibration curves, and decision curves, with SHapley Additive exPlanations (SHAP) analysis employed for interpretability.

The inclusion of peritumoral features improved the diagnostic performance to varying degrees, with the model incorporating a 10 mm peritumoral region achieving the highest overall accuracy. Combining radiomics with clinical features further enhanced predictive performance. The LGBM model outperformed the other algorithms across subgroups, achieving a maximum AUC of 0.909, an accuracy of 0.878, and an F1-score of 0.971. SHAP analysis revealed the contribution of key features, improving model interpretability.

This study demonstrates the value of integrating radiomics and clinical features for breast mass diagnosis, with optimized peritumoral regions enhancing model performance. The LGBM model emerged as the preferred algorithm due to its superior performance. These findings provide strong support for the clinical application of ABVS imaging and future multicenter studies, highlighting the importance of microenvironmental features in diagnosis.