AI-based triage and decision support in mammography and digital tomosynthesis for breast cancer screening: a paired, noninferiority trial.

This was a prospective, paired, noninferiority clinical trial assessing whether an AI system could safely reduce radiologist workload in breast cancer screening by autonomously excluding low-risk mammograms from human review.

Design and population

• Timeframe: March 2022 – January 2024
• Setting: Routine screening program
• Participants: 31,301 women undergoing screening mammography
• Parallel strategies applied to all exams:
  1. Standard of care: double-blind human reading
  2. AI-supported strategy:
     • AI classification: exams classified as low or non–low risk
     • Low-risk exams: automatically assessed as normal (no radiologist review)
     • Non–low-risk exams: double reading with AI support

Primary outcomes

• Radiologist workload
• Cancer detection rate (CDR)
• Recall rate

Key results

• Workload:
  • AI strategy: reduced radiologist workload by 63.6% (nearly two-thirds fewer human reads)
• Cancer detection:
  • CDR increase: 15.2% (95% CI 6.6%–24.4%)
  • Absolute increase: from 6.3/1,000 to 7.3/1,000 screens (P < 0.001)
• Recall rate:
  • Noninferiority: not met
  • Change with AI: recall rate 14.8% higher (95% CI 9.0%–20.6%)

Modality subanalyses

• Digital mammography (DM):
  • Workload reduction: −62.1%
  • CDR change: increased by 1.6/1,000
  • Recall rate change: increased by 1.3 percentage points
• Digital breast tomosynthesis (DBT):
  • Workload reduction: −65.5%
  • CDR: remained stable versus standard reading
  • Recall rate: remained stable versus standard reading

Interpretation for practice

• Feasibility and impact: A partially automated AI workflow, in which low-risk exams bypass human reading, is feasible and substantially decreases radiologist workload while modestly improving cancer detection overall.
• Trade-offs: The trade-off is an increased recall rate, particularly for DM, whereas DBT maintained recall and detection performance with similar workload savings.
• Implications: These data support integration of AI triage into screening programs, especially where workforce constraints are significant, but highlight the need to monitor and manage recall rates.