Fusion of multimodal ultrasound radiomics and deep learning for differentiating benign and malignant breast masses

Background Differentiating benign from malignant breast masses remains challenging because grayscale morphology, vascularity, and tissue stiffness provide complementary information that is not fully captured by a single ultrasound modality. This study evaluated a diagnostic framework integrating multimodal ultrasound radiomics and deep learning. Methods A single-center retrospective patient-level dataset of pathologically classified breast masses included B-mode ultrasonography, color Doppler flow imaging (CDFI), and shear wave elastography (SWE) with shear wave velocity (SWV) measurements. Clinical, conventional ultrasound, and quantitative SWV features were analyzed. Clinical, B-mode radiomics, multimodal radiomics, deep learning, and combined models were developed in a patient-level training set and evaluated in an internal test set. The internal hold-out test set comprised 135 patients, including 54 malignant and 81 benign lesions (40.0% malignant). Fivefold out-of-fold (OOF) predictions were generated within the training set for internal validation. Results Malignant lesions occurred in older patients, were larger, and showed more suspicious morphology, vascularity, and higher SWV than benign lesions.Fivefold OOF AUCs were 0.721, 0.842, 0.892, 0.920, and 0.928. In the internal test set, the respective AUCs were 0.768, 0.833, 0.901, 0.930, and 0.941. The combined model had the highest observed AUC (0.941, 95% CI 0.900-0.976) and the highest specificity (0.889); its sensitivity, accuracy, PPV, and NPV were 0.870, 0.881, 0.839, and 0.911, respectively. The combined model did not differ significantly from multimodal radiomics (Holm-adjusted P = 0.456) or deep learning (P = 0.698). Conclusion Score-level fusion of clinical information, multimodal ultrasound radiomics, and deep-learning features was evaluated for differentiating benign and malignant breast masses. The combined approach showed the highest numerical AUC with nonsignificant differences relative to multimodal radiomics or deep learning after multiplicity adjustment.