MBUCF: a cross-fusion method for multimodal breast ultrasound based on multidimensional features.

Background: Breast cancer (BC), characterized by high incidence and mortality rates, poses a severe threat to the life and health of women worldwide. Early, accurate identification is crucial for improving patient prognosis. Existing single-modal ultrasound techniques are insufficient to meet clinical needs. This study aims to establish a high-performance, lightweight multimodal ultrasound fusion model, providing a feasible method for distinguishing benign from malignant breast tumors. Methods: Conventional ultrasound and elastography images of 227 patients with breast tumors were collected from Tianjin Cancer Institute and Hospital. After image registration, region of interest (ROI) annotation, and feature extraction, 581-dimensional features were extracted. A feature selection method of variance rough screening-information rough screening-sparse regression was adopted to select 10% of core features to achieve model lightweighting. Two models, light gradient boosting machine (LightGBM) and extreme gradient boosting (XGBoost), were selected. Scale_pos_weight and synthetic minority oversampling technique (SMOTE) were used to handle class imbalance, and relevant parameters were set empirically. In this study, metrics such as area under the curve (AUC), balanced accuracy (balanced Acc), and macro F1-score (F1-macro) were used to evaluate the performance of the model. Results: The proposed feature_cross fusion method in this study achieved the optimal performance. Under the full‑feature setting, the optimal model yielded an AUC of 0.8655 [95% confidence interval (CI): 0.7518-0.9539], a balanced Acc of 0.7652, and an F1‑macro of 0.7598. When only 10% of the features were retained, the model still maintained excellent performance with an AUC of 0.8258 (95% CI: 0.7083-0.9365), a balanced Acc of 0.7614, and an F1‑macro of 0.7608, and effectively achieved model lightweighting. Conclusions: The multimodal ultrasound model based on feature cross-fusion can effectively improve the accuracy of distinguishing benign from malignant breast tumors and meet the lightweight requirement, providing a feasible tool for the early non-invasive diagnosis of BC.