An artificial intelligence system developed by Google DeepMind and the Royal Marsden NHS Foundation Trust has achieved 94.5% accuracy in detecting breast cancer on screening mammograms, outperforming the average radiologist accuracy by 11.5 percentage points in a prospective clinical study. Results published in Nature Medicine in March 2026 are fueling a real debate about AI's role in clinical radiology.
How the system works
Trained on over 2.8 million annotated mammography images, the model uses a deep convolutional neural network architecture that simultaneously analyzes lesion morphology, breast tissue density, and calcification distribution patterns. Unlike earlier systems, it integrates patient clinical data — age, family history, BMI — to generate an individualized risk score.
The full analysis takes under 40 seconds per study, compared to the 3-8 minutes required by a radiologist under standard conditions. For a service processing 500 mammograms daily, the operational efficiency gains are substantial.
Key clinical results
The study evaluated the system on a prospective cohort of 28,000 women screened within the NHS between 2023 and 2025. The most relevant findings:
- Sensitivity of 94.5% versus 83% for standard double reading.
- 37% reduction in false positives, meaning fewer unnecessary biopsies and less patient anxiety.
- Improved early detection: the system identified stage I lesions in 78% of positive cases, versus 61% with conventional protocol.
- 5.5% false negative rate, which researchers highlight as an area for improvement before any autonomous clinical deployment.
The model showed stronger performance on dense breast tissue (BI-RADS categories C and D), precisely where human reading is most challenging and where the highest proportion of interval cancers occur.
The diagnostic autonomy debate
The publication has generated divided positions in the radiology community. The European Society of Radiology (ESR) stated that results are "promising but insufficient to justify autonomous AI reading without medical supervision." The main arguments against rapid deployment include:
- The model was trained and validated on a predominantly Caucasian UK population, with potential performance bias in other ethnic groups.
- The 5.5% false negative rate, while below the human average in this study, can have serious clinical consequences at scale.
- Legal liability in case of diagnostic error remains unresolved under any European regulatory framework.
Proponents of faster integration argue the debate should not frame AI versus radiologist, but rather how AI can relieve the burden on healthcare systems where radiologist shortages are already a structural problem. The UK, for instance, had over 1,600 unfilled radiology posts as of 2025.
Regulatory pathway and next steps
In the European Union, AI-based imaging diagnostic systems are classified as class IIb or III medical devices under the Medical Device Regulation (MDR 2017/745), requiring clinical trials and notified body oversight before any autonomous clinical use.
The developing consortium has begun discussions with the EMA and the UK's MHRA to design a multiethnic validation protocol. The goal is to have a certified version for assisted use — where the radiologist makes the final call supported by the AI report — operational in NHS pilot centers by Q4 2026.
