Multimodal feature fusion model for breast mass malignant risk stratification.

Introduction: This study aims to develop and validate machine learning models that integrate multimodal features (BI-RADS terminology, ultrasound imaging, and radiomics) to improve breast mass malignancy risk stratification and compare their diagnostic performance across different BI-RADS categories. Methods: This retrospective cohort study analyzed data from 2, 685 patients with 3, 703 ultrasound images collected from July 2019 to March 2024 at a single medical center. Patients included women with complete ultrasound images and clear pathological diagnoses. The dataset comprised 2, 069 benign cases (2, 762 images) and 616 malignant cases (941 images), randomly divided into training (n=2, 979 images) and validation (n=724 images) sets. Primary outcomes were diagnostic accuracy and area under the receiver operating characteristic curve (AUC) for distinguishing malignant from benign breast masses. Three machine learning models (Logistic Regression, Support Vector Machine, and Random Forest) were trained using BI-RADS terminology features, ultrasound quantitative features, radiomics features, and combined multimodal features. Performance was evaluated both overall and within specific BI-RADS subcategories (2, 3, 4a, 4b, 4c, and 5). Results: Among 2, 685 patients, the Random Forest model using combined multimodal features achieved the highest overall performance with an AUC of 0.850 (95% CI 0.810- 0.875). For single-modality approaches, Logistic Regression performed best with BI-RADS terminology features, with an AUC of 0.820 (95% CI, 0.775-0.856), and radiomics features, with an AUC of 0.740 (95% CI, 0.706-0.780); while Random Forest was optimal for ultrasound imaging features, with an AUC of 0.800 (95% CI, 0.768-0.839). Subgroup analysis revealed excellent performance for BI-RADS categories 2 (AUC, 1.000-1.000) and 3 (AUC, 0.947-0.957), acceptable performance for 5 (AUC, 0.813-0.870) and 4a (AUC, 0.800-0.867), but poor performance for categories 4b (AUC, 0.649-0.709), 4c (AUC, 0.551-0.623). Discussion: This study demonstrates that machine learning models integrating multimodal ultrasound features can effectively stratify breast mass malignancy risk, with the Random Forest model using combined features showing superior performance. The approach shows particular strength in BI-RADS categories 2, 3, 5 and 4a, suggesting potential clinical utility for reducing unnecessary biopsies and improving diagnostic confidence. However, performance limitations in higher-risk categories (4b, 4c) indicate need for further model refinement and multicenter validation before clinical implementation.