AI Rivals Radiologists in Breast Cancer Detection, But Challenges Remain for NHS Adoption

Artificial intelligence is demonstrating the potential to match, and in some cases exceed, the performance of human radiologists in detecting breast cancer, offering a potential solution to growing workloads and staffing shortages within the National Health Service (NHS). However, a recent study highlights that integrating this technology isn’t as straightforward as simply replacing human expertise.

AI’s Performance: A Closer Look

New research, published in Nature Cancer, analyzed over 175,000 breast scans from five NHS screening services. The study, conducted by Imperial College London, Google, the universities of Cambridge and Surrey, and several NHS Trusts including Imperial College Healthcare, Cambridge University Hospitals, and The Royal Marsden, found that AI detected more cases of invasive cancer and overall cancer cases, while simultaneously reducing false positives and the number of women recalled for additional screening after an initial scan.

Specifically, the AI detected around two more cancers per 1,000 women screened compared to human radiologists. Importantly, it also identified 25% of “interval cancers” – cancers diagnosed between routine screenings after a previous scan appeared clear – suggesting the potential for earlier detection and improved patient outcomes.

The Human-AI Partnership: Not a Replacement

While AI outperformed a single radiologist, the NHS typically employs a “double-reading” system where two radiologists independently review each scan, with a third expert arbitrating any disagreements. When comparing a human-AI team to a human-human pair, the researchers found results were largely equivalent. The AI showed a slight advantage in identifying difficult-to-detect cancers but also flagged more cases that ultimately proved to be benign.

As the researchers concluded, AI operates “at least on par with human specialists.” This suggests AI isn’t poised to replace radiologists, but rather to augment their capabilities, offering a valuable second opinion and potentially accelerating the diagnostic process.

Time Savings and the Radiologist Shortfall

The study revealed a significant time saving with AI assistance. The average time for AI to complete a scan read was 17.7 minutes, compared to 2.08 days for the first human radiologist. This efficiency is particularly crucial given the current and projected shortage of clinical radiologists in the UK. The Google paper notes a current 30% shortfall, predicted to rise to 40% by 2028. The Royal College of Radiologists also highlights the growing workforce challenges.

Challenges to Implementation: Arbitration and Calibration

Despite the potential benefits, the study revealed complexities in implementation. The use of AI increased the need for arbitration – a third expert review – by 142% and 22% at the two centers studied. Human doctors found it more challenging to trust the AI’s evaluations, leading to more cases requiring expert adjudication. In 93 instances, the AI correctly identified cancer, but human radiologists overruled the assessment during arbitration, often due to uncertainty about the AI’s reasoning.

the AI’s performance was sensitive to changes in equipment. When radiologists switched machines, recall rates for patients doubled, indicating the AI required careful calibration to each local environment. The researchers emphasize the need for a “phased, iterative approach to AI deployment” to ensure accurate and reliable results.

The Future of AI in Breast Cancer Screening

While the path to widespread AI adoption in the NHS isn’t without hurdles, the potential benefits are substantial. The technology offers a promising avenue for improving diagnostic accuracy, reducing radiologist workload, and saving lives. However, successful integration requires careful planning, ongoing monitoring, and a commitment to continuous calibration and refinement. The NHS must adapt its infrastructure and workflows to effectively leverage the power of AI while maintaining patient safety and trust.