Google’s artificial intelligence to fight breast cancer: encouraging results
In a study published in Nature, the effectiveness of the use of artificial intelligence to support the work of doctors and radiologists in identifying cases of breast cancer is shown.
Google is actively supporting the development of artificial intelligence that can help doctors better identify cases of breast cancer, currently the most common in the world, along with lung cancer.
Early diagnosis is the best weapon in combating this pathology, and today the most common tool for carrying out the appropriate analyzes is the mammograph, which suffers from some significant problem with false positives and false negatives and has an error rate of about 20% all in all.
In a study funded by Google and the results of which were published in the latest edition of Nature and illustrated on the official blog of the Mountain View company, the researchers used an information database consisting of 28 thousand mammograms: 25 thousand of these of women from the United Kingdom, while the remaining 3,000 US women.
The researchers trained AI in examining x-ray images of mammograms and later to identify signs of cancer by observing the breast changes in the sample. The AI assumptions were then compared with the outcome of the real cases.
In this way, the system was able to reduce false negatives by 9.4% and false positives by 5.7% for the US sample. As for the UK sample, the model reduced the false negatives by 2.7% and the false positives by 1.2%; this is because normally in the United Kingdom, two radiologists check mammography images.
Although the system has shown the ability to achieve a more accurate result overall, there has been no shortage of cases where doctors have successfully identified a tumor where AI has failed to detect it.
And this is precisely the reason why Google frames this project as a resource that can be added, in clinical practice, to radiologists to help them improve diagnoses. Both have their strengths, which combined can lead to an improvement in overall results.
Researchers are now working to understand if the functioning of the model can be generalized towards the wider population because obviously, there is always some difference at the moment in which such a system is applied to clinical practice in the real world.
