Diagnostic accuracy of artificial intelligence-assisted early detection of breast cancer using ultrasound imaging: A systematic review
Artificial intelligence (AI) has emerged as a promising tool for improving the diagnostic accuracy of breast ultrasound and facilitating early breast cancer detection. This systematic review evaluated the current evidence on AI-based diagnostic models for breast ultrasound imaging. A systematic literature search was performed in PubMed, Scilit, and PMC following the PRISMA 2020 guidelines. A total of 1,296 records were identified, and after duplicate removal and eligibility screening, 18 studies published between 2021 and 2026 were included. Methodological quality was assessed using the QUADAS-3 tool. Owing to heterogeneity in AI models, study designs, and reported outcomes, a qualitative synthesis was conducted without quantitative meta-analysis. AI-assisted breast ultrasound demonstrated generally favourable diagnostic performance, although the reported parameters varied across studies. Sensitivity ranged from 71.4% to 100.0%, specificity from 71.6% to 96.2%, and AUC from 0.778 to 1.00. One EfficientNet-B7 model integrated with explainable AI achieved an AUC of 1.00, sensitivity of 99.5%, an F1-score of 98.9%, and accuracy of 99.14%, representing one of the highest-performing models identified. Other high-performing approaches included hybrid deep learning, Vision Transformers, interpretable ensemble transformers, convolutional neural networks, and machine learning models, which improved lesion classification, reduced false-positive findings, enhanced interpretability, and supported standardized reporting. However, most studies were retrospective, single-center investigations with limited external validation. AI-assisted breast ultrasound is a promising adjunctive tool for early breast cancer detection. Although current evidence supports its clinical potential, further prospective multicenter studies with standardized methodologies and external validation are needed before routine clinical implementation.