New studies have tested how Google's artificial intelligence system for breast cancer detection performs in real-world settings, analyzing mammograms in British hospitals. Although this system was introduced in 2020, it has not yet been routinely used for patient diagnosis.

The research shows that AI can detect more tumors in some situations than a human expert, while also reducing the number of false positives.

How the system works

Google's system uses three convolutional neural networks that have been trained on extensive databases of mammographic images. These models work together to create an image representation, identify suspicious areas, and then assess the probability of cancer.

The development and testing involved scientists from Imperial College London, the University of Surrey, the Royal Surrey NHS Foundation Trust, and other screening centers within the National Health Service.

Test results

In a retrospective study involving 116,000 mammograms of women aged 50 to 70, the system achieved a sensitivity of 0.541, while the initial human assessment achieved 0.437. This means that the AI detected more actual cases of cancer.

At the same time, it maintained a high specificity of 0.943, which is only slightly lower than the 0.952 achieved by doctors, but statistically comparable.

A significant finding was that the system was able to identify approximately 25% of cases that were initially missed by human doctors, but in which cancer was later confirmed in the patients.

In a simulation where the AI replaced the second reader in a double-reading process, it showed slightly better results than a human, and also had the potential to reduce the workload of specialists by approximately 40% due to more efficient case triage.

Real-world testing

In another study, approximately 9,250 new images from 12 British clinics were analyzed in 2023 and 2024. The AI system operated in parallel with standard diagnostics, but did not influence the decisions of the doctors.

Here, it was shown that the AI was significantly faster: the average evaluation time for an image was 17.7 minutes, compared to more than two days for the initial human assessment.

Even in this test, the system maintained a higher sensitivity than the initial medical assessment and a comparable specificity.

Benefits and limitations

The results confirm that AI can significantly help to alleviate the burden on healthcare systems, especially in areas where there is a shortage of radiologists and a high volume of screenings.

However, some doctors expressed distrust in the system's outputs, which highlights that the key issue is not only technical accuracy, but also the acceptance of the technology in clinical practice.

Broader context

The use of artificial intelligence in mammography builds on developments dating back to the 1990s, when the first computer-assisted diagnostic systems emerged. A significant leap forward occurred with deep learning in the mid-2010s, when models began to outperform traditional methods.

Breast cancer affects approximately 2.3 million women worldwide each year, and hundreds of thousands of cases result in death. Early diagnosis remains crucial.

deeplearning.ai/gnews.cz - GH