Could AI help the NHS detect breast cancer earlier?

Introduction

Breast cancer (BC) is the most common cancer among women in the UK, with approximately 59,400 new cases diagnosed annually and around 1 in 8 women likely to develop BC during their lifetime. Early detection through screening and timely diagnosis can help reduce mortality rates and improve treatment outcomes. While mammography is the standard method for BC screening (Figure 1), it is not without limitations, as some BC cases may go undetected and subsequently be diagnosed clinically within one year of screening. 

Current NHS breast screening practices

The NHS Breast Screening Programme (NHS BSP) routinely invites women aged 50-70 for mammographic screening, while those aged 71 and over can self-refer. The mammograms are independently examined by two trained readers, including radiologists, breast clinicians and trained radiographers. If the readers disagree, arbitration by a third reader or group consensus may be used to reach a final decision. The expected waiting time for screening results is 2-3 weeks, which may be a source of anxiety for some women during this period. The UK has also reported a 30% shortfall of clinical radiologists, which was projected to increase to 40% by 2028, raising concerns about the long-term sustainability of the NHS BSP.

Artificial Intelligence (AI) in BC screening

AI-assisted screening could help improve BC detection while reducing the workload of radiologists within the NHS BSP (Figure 2). Previous research has shown that standalone AI tools can perform as well as or better than radiologists in detecting BC. Despite this, the use of AI algorithms has not been widely implemented, and their impact when integrated into clinical screening workflows is less well understood.

To investigate this, Kelly et al. (2026) evaluated Google’s AI mammography system using 115,973 mammograms from 5 NHS breast screening services, assessing its diagnostic performance against human readers and its potential to replace one human reader in the existing screening process. 

The study consisted of two phases. The first was a retrospective phase, where researchers analysed previously collected mammograms to compare AI performance with human readers. This was followed by a prospective observational phase across 12 NHS screening sites to investigate how the AI system could be integrated into routine BC screening.

The retrospective phase

In the retrospective phase, the researchers compared the AI system’s ability to detect BC with that of human readers, using sensitivity and specificity to assess its accuracy. They also used up to 39 months of existing patient follow-up records to identify cancers diagnosed after the initial screening.

The findings showed that Google’s AI system achieved higher case-level sensitivity than the first human reader (54.1% vs 43.7%), suggesting it correctly identified a greater proportion of BC cases. However, its specificity was slightly lower (94.3% vs 95.2%), indicating that AI was more likely to incorrectly flag women without BC (Figure 3). Despite this difference, the researchers found that AI’s specificity was not significantly worse than the first human reader’s, remaining within the predefined acceptable margin.

Moreover, AI retrospectively identified 25% of interval cancer cases, which are cancers diagnosed after a normal screening result but before the next scheduled screening appointment. It also identified 25.1% of BC cases that were subsequently detected at the next screening appointment, which occurred approximately 3 years later. These findings suggest that AI could help detect some BC cases earlier, although further research is needed to establish whether AI would lead to earlier diagnoses in clinical practice. 

The researchers also estimated that by replacing one human reader with AI, it could reduce total human reading time by 32.1%, although the number of arbitration reads required was estimated to increase by 60.3%.

The prospective phase

During the prospective observational phase, Google’s AI system analysed mammograms with the two human readers involved in routine NHS screening. However, the AI’s predictions were recorded separately and did not influence clinical decisions or patient care.

The findings showed that AI maintained a similar ability to distinguish between BC and non-cancer cases compared with the retrospective phase, suggesting that its diagnostic accuracy could be maintained when applied to newly collected NHS screening data. However, differences between historical and newer screening data, known as distribution shift, meant that the AI system initially flagged more women for further examination than expected, raising concerns about false-positive results. The researchers therefore adjusted the AI system’s decision threshold, which reduced the recall rates, indicating that fewer mammograms were flagged as suspected cases. These findings highlight the importance of monitoring and adjusting AI systems to ensure reliable performance in clinical practice.

The median time from screening to a completed AI read was 17.7 minutes, compared with 2.08 days for the first human read. This suggests that AI could help speed up the process of mammogram analysis, although the study did not determine whether this would reduce waiting times for patients in routine clinical practice.

Challenges and costs

The researchers also identified practical challenges to implementing AI into routine NHS breast screening. Existing NHS screening software could not automatically incorporate AI predictions, with 8 of the 9 screening services interviewed still relying on paper-based processes. These findings suggest that introducing AI would require changes to existing workflows and investment in digital infrastructure. 

The long-term adoption of AI also depends on women’s willingness to accept its use. Gatting et al. (2026) surveyed 3,419 women in England and found that using AI together with a human reader to examine existing mammograms was more acceptable than using AI to determine which mammograms required human review. This highlights the importance of considering public attitudes towards AI to avoid worsening existing inequalities in breast screening. 

The financial implications must also be considered. A recent economic modelling study using a different AI system estimated that replacing one human reader with AI could save approximately £31 per woman invited for screening over a lifetime, while increasing BC detection and reducing BC-related deaths. However, these findings were based on simulations rather than real-world NHS implementations, highlighting the need for further economic evaluations.

Conclusion

To conclude, Kelly et al. (2026) suggest that AI could improve BC detection and reduce the workload of NHS screening professionals. However, successful implementation would require careful integration into existing NHS screening practices, consideration of costs, ongoing monitoring of AI performance and full transparency into how the system generates its predictions. These measures could help build trust among clinicians while supporting informed clinical decision-making. 

Written by Caroline Valentine

Related articles: Evolution of AI / Breast cancer in males / Revolutionising patient setup in cancer treatment / AI in diagnosis and treatment of genetic diseases

REFERENCES

Breast cancer incidence (invasive) statistics | Cancer Research UK [Internet]. [cited 2026 Oct 14]. Available from: https://www.cancerresearchuk.org/health-professional/cancer-statistics/statistics-by-cancer-type/breast-cancer/incidence-invasive

Breast cancer awareness: Surrey and Sussex Healthcare NHS Trust [Internet]. www.surreyandsussex.nhs.uk. [cited 2026 Oct 17]. Available from: https://www.surreyandsussex.nhs.uk/our-services/cancer-services/breast-unit/breast-cancer-awareness

Shi J, Li J, Gao Y, Chen W, Zhao L, Li N, et al. The screening value of mammography for breast cancer: an overview of 28 systematic reviews with evidence mapping. J Cancer Res Clin Oncol. 2025 Mar 6;151(3):102. doi:10.1007/s00432-025-06122-z 

Cancer Research UK. Mammogram [Internet]. www.cancerresearchuk.org. 2023 [cited 2026 Oct 17]. Available from: https://www.cancerresearchuk.org/about-cancer/tests-and-scans/mammogram

Reeves RA, Kaufman T. Screening Mammography in Breast Cancer Detection. In: StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; 2023 [cited 2026 Oct 14]. Available from: http://www.ncbi.nlm.nih.gov/books/NBK559310/ 

Cornford E, Cheung S, Press M, Kearins O, Taylor-Phillips S. Optimum screening mammography reading volumes: evidence from the NHS Breast Screening Programme. Eur Radiol. 2021 Sep 1;31(9):6909–15. doi:10.1007/s00330-021-07754-8

Warren LM, Venton J, Young KC, Halling-Brown M, Kelly CJ, Wilson M, et al. Impact of using artificial intelligence as a second reader in breast screening including arbitration. Nat Cancer. 2026 Mar;7(3):507–21. doi:10.1038/s43018-026-01128-z

Gatting L, Jones CK, Jamshidi B, Kehagia AA, Waller J. Acceptability of using artificial intelligence in the National Health Service breast screening program: A randomized online survey of screening-eligible women in England. Mayo Clinic Proceedings: Digital Health [Internet]. 2026 Mar [cited 2026 Oct 17];4(1):100329. Available from: https://www.mcpdigitalhealth.org/article/S2949-7612(25)00136-1/fulltext

Kelly CJ, Wilson M, Warren LM, Sidebottom R, Halling-Brown M, Yang L, et al. Diagnostic accuracy, fairness and clinical implementation of AI for breast cancer screening: results of multicenter retrospective and prospective technical feasibility studies. Nat Cancer. 2026 Mar;7(3):494–506. doi:10.1038/s43018-026-01127-0

Shieh Y, Ziv E, Kerlikowske K. Interval breast cancers — insights into a complex phenotype. Nat Rev Clin Oncol. 2020 Mar;17(3):138–9. doi:10.1038/s41571-020-0327-9 

Hill H, Roadevin C. Economic evaluation of artificial intelligence for cancer detection in the UK breast screening programme. Br J Cancer. 2026 Aug 5;135(3):453–60. doi:10.1038/s41416-026-03465-3 

Pesapane F, Rotili A, Penco S, Dominelli V, Priolo F, Marinucci I, et al. Explainable Artificial Intelligence in Mammography: A Systematic Review of Methods, Evaluation Practices, and Clinical Readiness. Diagnostics (Basel). 2026 May 6;16(9):1412. doi:10.3390/diagnostics160914s