Artificial intelligence in breast cancer screening: A systematic review and meta-analysis of integration strategies.

Authors

Abstract

To compare AI-augmented and conventional double reading in organised breast-cancer screening with respect to cancer-detection rate (CDR), recall rate, and radiologist workload. We conducted a systematic review and random-effects meta-analysis of 13 prospective and retrospective studies (1.03 million screens) from 2017 to 2024 that embedded commercial or research AI into population-based digital mammography or tomosynthesis programmes. Eligible studies included ≥ 10,000 screens (or ≥100 cancers) and reported CDR, recalls, and/or workload metrics. We extracted cancer and recall counts and calculated risk ratios (RRs) for AI-augmented versus double reading, overall and by integration model: independent second reader, gate-keeper/decision-referral triage, and concurrent overlay. Overall, AI-augmented protocols achieved CDR parity (RR 1.01; 95 % CI 0.96-1.07) and no significant change in recalls (RR 1.00; 95 % CI 0.88-1.15). Triage models preserved CDR (RR 1.02; 95 % CI 0.98-1.07) while reducing recalls by 11 % (RR 0.89; 95 % CI 0.82-0.96) and cutting initial reads by 44-70 %. Independent-reader workflows maintained CDR (RR 0.98; 95 % CI 0.92-1.05) but showed variable recall effects (RR 1.12; 95 % CI 0.90-1.39) driven by arbitration logic and threshold choices. Concurrent overlay (two studies) indicated possible sensitivity gains (RR 1.31; 95 % CI 0.90-1.91) without higher recall rates, though precision was limited. AI integration can match conventional double reading in detection performance, but its impact on workflow depends on the chosen model. Triage-based approaches consistently lower radiologist workload and recalls without compromising sensitivity, whereas replacing a second reader may simply shift effort to arbitration. Future implementation should focus on workflow-aware metrics and prospective threshold validation.

Tags

Topics