Evidence explorer
Read the study by layer.
Move between the research question, method, recorded result, and interpretation boundary without losing context.
Artificial intelligence (AI) for breast cancer detection is moving from retrospective benchmarks toward prospective workflow evaluation. This narrative review synthesizes evidence across mammography and digital breast tomosynthesis, ultrasound, magnetic resonance imaging, histopathology, thermography, and clinically coherent multimodal systems. A structured narrative search screened 730 Scopus-indexed records published from January 2023 to 20 July 2026; 99 indexed records plus 21 additional sources formed a 120-source core, and 13 targeted sources broadened annotation-efficient learning, multimodal foundation models, regulation, cybersecurity, and health economics, yielding 133 cited sources. Evidence was weighted by clinical alignment, cohort provenance, reference-standard quality, external validation, statistical uncertainty, and implementation relevance. Prospective mammography studies show that AI-supported reading can increase cancer detection and reduce workload, but effects depend on workflow and may alter recall. Other modalities remain dominated by retrospective studies despite advances in transformers, one-stage detectors, self-, semi-, weakly supervised, and few-shot learning, graph models, and multimodal fusion. Reported accuracies above 95 % are therefore less informative than external discrimination, calibration, subgroup failure analysis, and patient-pathway outcomes. We identify a translational triad—label efficiency, cross-domain generalizability, and workflow-aware trustworthiness—and propose an evidence hierarchy and hypothesis-generating functions for label efficiency, generalization gap, workflow utility, incremental multimodal value, and translational sufficiency. These constructs require prospective validation. Priorities are patient-level separation, multi-institutional validation, density- and vendor-stratified calibration, clinically coherent fusion, tested explanations, lifecycle governance, cybersecurity, economic evaluation, and post-deployment monitoring. AI should be developed as a monitored component of the detection pathway rather than as an autonomous image classifier.