A research recently published in Nature reports the results of an AI model developed by Google for the diagnosis of breast cancer by scanning digital mammograms.
Breast cancer is the second leading cause of death in women. Early diagnosis is the best form of prevention for the identification and treatment of the disease. Mammograms are an effective diagnostic test, but they do not identify about 20% of breast cancers and also present a significant problem with false negatives and false positives.
The research was carried out by Google and DeepMind, in collaboration with Cancer Research UK Imperial Centre, Northwestern University and Royal Surrey County Hospital, to assess whether artificial intelligence can help radiologists more accurately detect signs of breast cancer.
The model was trained and developed on a representative data set that included anonymous mammograms of over 76,000 women in the UK and over 15,000 women in the US. The model was then evaluated on an anonymous data set of over 25,000 women in the UK and over 3,000 women in the US. In this evaluation, the system produced a 5.7% reduction in false positives in the US, and a 1.2% reduction in the UK.
The researchers then wanted to see if the model could be extended to other healthcare systems. They then trained the model only on women’s data in the UK and then evaluated it on the women’s data set in the US. In this separate experiment, there was a 3.5% reduction in false positives and an 8.1% reduction in false negatives, showing the model’s potential to generalize to new clinical contexts while continuing to operate at a more precise level than radiologists.
In particular, in making its decisions, the model received less information than radiologists. The radiologists (in line with routine practice) had access to previous patient histories and mammograms, while the model only processed the most recent anonymous mammography without further information. Despite the use of these X-ray images alone, the model outperformed individual radiologists in precisely identifying breast cancer.
However, the system is not perfect. While researchers found that AI outperformed doctors in identifying breast cancer in most cases, there were also cases where doctors reported cancer that the model did not report.
Google presents this model as a tool to help radiologists, not to replace them. To date the capabilities of both are complementary, there are a number of cases where radiologists draw on the model, and vice versa. By integrating the two resources, overall results could be improved.