AI for mammologists
Doctors have taught artificial intelligence to accurately predict breast cancer
Doctors from the United States have created an artificial intelligence that can very accurately assess the likelihood of developing breast cancer in a woman, relying only on X-rays of her mammary glands. The results of the first "field trials" of this program were presented in the journal Radiology (Yala et al., A Deep Learning Mammography-based Model for Improved Breast Cancer Risk Prediction).
"Mammograms not only allow us to assess the degree of breast density, but many other hints of cancer development are hidden there, invisible to us. The use of deep learning and artificial intelligence allows us to "pull out" these signs and give an accurate assessment," said Adam Yala from the Massachusetts Institute of Technology in Cambridge (in a press release Using AI to predict breast cancer and personalize care – VM).
To date, breast cancer remains the most common type of malignant tumors among women. According to WHO estimates, about 500 thousand women die from it every year. The problem with the treatment of breast cancer is that the detection of breast cancer in the early stages of development is difficult due to the imperfection of diagnostic methods - mammography correctly detects it only in 12% of cases.
Genetic studies in recent years show that breast tumors most often develop in women who are carriers of mutations in a small set of genes, such as HER2, BRCA, ESR1 and a number of other DNA sites responsible for repairing damaged chromosomes and "self-destruction" of cells in case of fatal genome breakdown.
The presence of these mutations in itself does not guarantee that a tumor will necessarily occur in a woman's body, and doctors usually rely on forecasts not only on genetic data, but also on the history of diseases in the family and mammograms, X-rays of the mammary glands. These data make it possible to significantly improve the accuracy of forecasts, but it still remains quite low.
Yala and his colleagues solved this problem by using the increasingly popular deep learning algorithms used to create the most advanced artificial intelligence systems. Recently, doctors have become interested in them, having discovered that in some cases such programs more often find traces of tumors in photographs and diagnose cancer more accurately than the most professional doctors do.
The main problem is that artificial intelligence needs to be trained for a long time and painstakingly, using data carefully selected and verified by humans. In the case of breast cancer, as the researchers note, this was not a problem due to its high prevalence and already existing tumor classification systems.
In total, scientists have collected over 90 thousand mammograms obtained during the examination of women in various clinics in the United States, as well as data on the incidence of cancer among them in the following months and years. This information, as well as family medical history and other indirect data, was used by doctors to train two AI systems, one of which took into account all the data in its work, and the second – only X-rays of the mammary glands.
As it turned out, artificial intelligence predicted the likelihood of developing cancer much more accurately than doctors, even if it had access only to mammograms – it correctly diagnosed it in 70% of cases, whereas doctors did it at best for 62% of patients.
The program is able to identify a woman at high risk of developing breast cancer four years (left) before it developed (right). The image is provided by the developers.
At the same time, the accuracy of his work did not decrease for representatives of various ethnic and racial minorities.
Now this algorithm, as Yala notes, has already been successfully used in the work of the Boston City Hospital, and in parallel, scientists are working to improve the accuracy of its predictions, teaching AI to use new types of data.
Portal "Eternal youth" http://vechnayamolodost.ru