Breast cancer is the most prevalent cancer among women. Early and accurate screening is crucial for improving patient outcomes. Ultrasound is a valuable diagnostic tool, particularly for dense breasts, yet its efficacy can be limited by operator dependency and interpretive variability. Artificial intelligence (AI) has shown significant potential to enhance the accuracy and efficiency of breast ultrasound. However, translating AI from research to clinical practice remains challenging due to several persistent gaps: the lack of robust clinical validation for generative AI in image enhancement; insufficient focus on AI for diagnosing non-mass lesions, which constitute a notable proportion of malignancies; and limited multi-center effectiveness data for commercial computer-aided diagnosis systems. This narrative review synthesizes recent advancements in AI for breast ultrasound and provides a critical, multifaceted analysis that integrates technological evolution, clinical-translation challenges, and implementation frameworks. Importantly, it highlights pervasive methodological limitations, such as small sample sizes, retrospective single-center designs, and inadequate external validation, that often lead to overestimation of real-world AI performance. By offering both actionable insights and a cautionary perspective, this review aims to guide the rigorous, evidence-based translation of AI into clinically viable tools.
Ma et al. (Thu,) studied this question.