Can artificial intelligence improve breast cancer screening without overburdening radiologists?

Breast cancer screening relies on the examination of mammograms by two healthcare professionals. While this method is effective, it faces a growing shortage of radiologists, particularly in the UK where nearly 40% of positions could be unfilled by 2028. A recent study explored the use of artificial intelligence as a second reader to analyze these images, aiming to reduce workload while maintaining diagnostic quality.

The study involved 50,000 women from two British screening centers. The results show that replacing the second human reader with an artificial intelligence tool maintains similar accuracy in terms of cancer detection and reduction of false positives. After an arbitration step, where experts re-examine disputed cases, the sensitivity and specificity of artificial intelligence proved to be as good as those of two human readers. This means the tool identifies as many real cancers and avoids as many incorrect diagnoses.

One of the main advantages of this approach is the significant reduction in workload. The number of mammograms requiring human analysis dropped by 50%, as artificial intelligence handles some of the examinations. However, arbitration becomes more frequent with artificial intelligence, increasing from 22% to 142% depending on the centers. This rise is explained by the fact that the tool can sometimes flag anomalies not confirmed by experts, or conversely, miss subtle cases that humans detect.

Artificial intelligence demonstrated a particular ability to identify cancers that appear between two screenings or during the next screening, known as interval or next-round cancers. Before arbitration, it detected these better than humans. But after re-evaluation by experts, this difference diminished, as some of the tool’s signals were dismissed. This raises a question: how can the suggestions of artificial intelligence be better integrated to make them more reliable and understandable for radiologists?

The tool’s performance varies depending on the equipment used for mammograms. It is more effective with devices from one manufacturer than another, suggesting it needs to be adapted to local technologies before deployment. Radiologists participating in the study expressed moderate confidence in the tool while pointing out its limitations, particularly its tendency to overestimate certain anomalies such as calcifications.

This study paves the way for broader use of artificial intelligence in breast cancer screening. It shows that it is possible to reduce the workload of professionals without sacrificing diagnostic quality. To move forward, the transparency of the tool’s decisions will need to be improved, and radiologists will need to be trained in its use so they can better evaluate its alerts and make the most of them.

Content References

Official Reference

DOI: https://doi.org/10.1038/s43018-026-01128-z

Title: Impact of using artificial intelligence as a second reader in breast screening including arbitration

Journal: Nature Cancer

Publisher: Springer Science and Business Media LLC

Authors: Lucy M. Warren; Jenny Venton; Kenneth C. Young; Mark Halling-Brown; Christopher J. Kelly; Marc Wilson; Megumi Morigami; Lisanne Khoo; Deborah Cunningham; Richard Sidebottom; Mamatha Reddy; Hema Purushothaman; Delara Khodabakhshi; Lesley Honeyfield; Amandeep Hujan; Tsvetina Stoycheva; Andy Joiner; Reena Chopra; Aminata Sy; Dominic Ward; Lin Yang; Rory Sayres; Daniel Golden; Namrata Malhotra; Rachita Mallya; Lihong Xi; Della Ogunleye; Charlotte Purdy; Alistair Mackenzie; Susan Thomas; Shravya Shetty; Fiona J. Gilbert; Ara Darzi; Hutan Ashrafian