Development and Evaluation of a Transformer-Based Multi-Scale Feature Model for the Classification of Benign and Malignant Breast Lesions in Ultrasound Image

This study introduces MSF-US, a novel deep learning framework integrating multi-scale feature fusion and attention mechanisms. It aims to enhance early breast cancer detection and improve diagnostic accuracy in ultrasound images, particularly addressing the challenges of capturing global and local features in dense breast tissue. The MSF-US model combines hierarchical multi-scale feature extraction with attention mechanisms to selectively focus on diagnostically relevant regions. The framework was rigorously validated on a multi-center dataset of breast ultrasound images. Its performance was compared against state-of-the-art models, such as ResNet-50, Vision Transformer (ViT), and Swin Transformer, evaluating accuracy, sensitivity, specificity, and area under the curve (AUC). In 5-fold cross-validation, MSF-US yielded an average accuracy of 0.9057, sensitivity of 0.9075, and AUC of 0.9212 with high net clinical benefit. On the independent C3 test set, it surpassed 8 SOTA baselines with a peak accuracy of 0.9406, F 1 score of 0.9483, and AUC of 0.9391. Evaluation on 3 diverse public cohorts (BUSI, BUS-BRA, and BrEaST) confirmed excellent cross-center generalizability, achieving AUCs up to 0.9519. Ablation analysis validated that the synergy of multi-scale feature fusion and attention mechanisms significantly boosted diagnostic precision over single-scale baselines. MSF-US delivers state-of-the-art performance in breast ultrasound imaging, offering a highly reliable tool for early breast cancer detection. Its innovative design successfully addresses the limitations of existing methods, presenting significant clinical potential to improve diagnostic accuracy and patient outcomes.