Background: Early detection of breast cancer and accurate assessment of lesions are key goals of imaging evaluation. Ultrasound is widely used, but its diagnostic performance is influenced by complex image features, noise, and operator experience. Reducing operator dependence and improving accuracy are critical clinical issues. Methods: In this retrospective study, 7,025 breast ultrasound images from our center were annotated based on pathology and split into training, validation, and internal test sets (8:1:1). The Dataset of Breast Ultrasound Images was used as the external test set. YOLO-v7 and YOLO-v8 models were trained through transfer learning after data augmentation and balancing the classes. Performance was compared on internal and external test sets and was evaluated against a reader study. Results: YOLO-v7 and YOLO-v8 reached optimal performance at epochs 294 and 135, respectively. YOLO-v7 slightly outperformed YOLO-v8 on the internal test set, while YOLO-v8 achieved higher accuracy, recall, specificity, precision, and F1 score on the external test set. Both models showed significantly higher accuracy, specificity, and precision than the senior radiologist, with YOLO-v8 achieving a significantly higher F1 score. Discussion: YOLO-v8 demonstrated better generalization due to its anchor-free mechanism and deeper architecture, while YOLO-v7 showed signs of overfitting. Both models outperformed the junior radiologist and approached or exceeded the diagnostic performance of the senior radiologist, indicating potential to assist less experienced readers. Conclusion: YOLO-v7 and YOLO-v8 effectively classified breast lesions. YOLO-v8 showed faster convergence and higher diagnostic efficiency, suggesting strong potential for clinical application.
Jiang et al. (Mon,) studied this question.