Breast Cancer Detection Algorithms: A Comparative Review of Leading Artificial Intelligence Models

Artificial intelligence (AI) has transitioned from proof-of-concept to clinical deployment in breast cancer screening, yet a comprehensive comparison of its performance across modalities and integration modes is needed. This comparative review followed PRISMA guidelines, synthesizing evidence from ۱۷ studies identified through systematic searches of PubMed, Scopus, and Web of Science (۲۰۱۷-۲۰۲۵). Included studies evaluated AI models on clinical breast imaging with pathology-confirmed outcomes, reporting clinically relevant endpoints like cancer detection rate (CDR), recall rate, and positive predictive value (PPV). Studies were critically appraised using QUADAS-۲ and ROBINS-I tools. AI’s clinical impact is contingent on its integration strategy. As an additional reader in double-reading, AI increased CDR by ۰.۷–۱.۶/۱۰۰۰ screens but raised workload by ۴–۶%. In single-read settings, AI-assisted radiologists increased CDR without elevating recall rates. Standalone AI at program scale showed strong discrimination (AUC ~۰.۹۳) but revealed trade-offs, such as lower PPV and under-detection of small tumors at sensitivity-matched thresholds. Multimodal systems (FFDM+DBT+synth۲D) demonstrated technical superiority for lesion localization and, in simulation, potential for substantial workload reduction (≈۴۴%) at fixed sensitivity. Prospective evidence for digital breast tomosynthesis (DBT) shows AI improves radiologist accuracy and reduces interpretation time. AI delivers measurable benefits in breast cancer screening, but its value is maximized when aligned with specific clinical workflows. Multimodal approaches show promise, but future work must prioritize prospective validation with standardized endpoints, interval cancer analysis, and comprehensive subgroup reporting to ensure generalizability and safety.

کلیدواژه ها:

Breast Neoplasms ، Artificial intelligence ، Early Detection of Cancer ، mammography ، algorithms