Breast cancer diagnosis could become more accurate with the help of deep learning algorithms, Google Health officials Shravya Shetty and Daniel Tse wrote in a company blog post this week. In fact, Google’s artificial intelligence for breast cancer diagnosis has been in the works for the past two years. And it looks like it’s already outperforming human radiologists.
How Google’s AI performed — False positives and false negatives can throw even the best of medical experts off and weaken their assessment of a patient who may or may not have breast cancer.
To see if the company’s artificial intelligence system could spot these complications, Google collaborated with Cancer Research United Kingdom Imperial Center, Northwestern University, and Royal Surrey County Hospital to initially train on de-identified mammograms from over 76,000 women in the United Kingdom and over 15,000 in the United States. Such mammograms have their identifiable data entirely removed.
The model was then run on de-identified mammograms from over 25,000 women from the United Kingdom and more than 3,000 women from the United States. In doing so, the Google model reduced false positives by 5.7 percent in the American and 1.2 percent in the United Kingdom data sets respectively. It also showed a reduction of false negatives by 9.4 percent in the American set and 2.7 percent in the United Kingdom's.
It’s a remarkable achievement; in spite of receiving mammograms only and no other information about the patients (like family history or prior tests), the system was able to accurately identify breast cancer more than human experts.
Will it be Dr. AI from now on? — In spite of the promising results that Shetty and Tse reported, it will take a long time before such a model becomes mainstream in medicine. As noted before, the system worked on anonymized mammograms only. Unlike human experts, it did not have a personal understanding of or access to patients’ family histories, past medical issues, and other crucial information that can point to the statistical likelihood of tumor growth and metastatic spread in a patient.
So, you don’t need to worry about your doctor being replaced by artificial intelligence. Google’s idea, it seems, is to just help clinicians spot these issues early on and improve patient care tenfold with the help of some algorithms in the mix.