Study Finds AI as Second Reader in Breast Screening Non-Inferior to Human Radiologists with Arbitration
Key Takeaways
- ▸AI as a second reader proved non-inferior to human radiologists (5% margin) in sensitivity and specificity when paired with arbitration protocols
- ▸Implementation of AI in breast screening could significantly reduce radiologist workload while maintaining diagnostic performance
- ▸Arbitration workflows both helped and hindered AI performance, overruling some incorrect recalls but also dismissing some valid cancer detections, indicating need for improved explainability
Summary
A large retrospective study of 50,000 women from two NHS breast-screening centers evaluated the impact of using artificial intelligence as a second reader in double-read breast-screening workflows. The research, which included arbitration protocols for conflicting readings, found that AI could replace the second human reader while maintaining non-inferior sensitivity and specificity compared to two human readers (P < 0.001), with the added benefit of reduced radiologist workload. The study examined 8,732 cases requiring arbitration, where human readers reviewed conflicting assessments between the first reader and the AI tool. While arbitration improved the AI arm's specificity by overruling incorrect AI recalls, it also occasionally overruled valid AI decisions for interval and next-round cancers, suggesting room for improvement in AI decision transparency and accuracy.
- Further development of AI tools with better explainability features could enable earlier cancer detection in breast screening programs
Editorial Opinion
This research provides compelling evidence that AI can effectively augment rather than replace human expertise in breast screening. The finding that arbitration sometimes overruled legitimate AI alerts underscores the importance of developing AI systems with greater interpretability—healthcare decisions require not just accuracy but also clinical transparency. The workload reduction benefits are significant for overburdened radiology departments, but this study wisely highlights that AI adoption must be coupled with continuous refinement of the technology and clear explainability mechanisms.
