This MIT AI Predicts Breast Cancer Risk Up to 5 Years in Advance

Breast cancer is among the top three causes of cancer-related death in women in the US, and while detection methods and technologies such as mammography have helped reduce the mortality rate by 39 percent since 1989, more than 41,000 women in the US will die of breast cancer this year, according to Cancer.net.

A new artificial intelligence model developed by researchers at MIT's Computer Science and Artificial Intelligence Labs (CSAIL), however, can analyze mammograms and predict breast cancer risks up to five years in advance.

Risk Models Are Flawed

Early detection is closely linked to reduced mortality rates, so research in the field has been centered on detecting symptoms as early as possible.

"Researchers have been creating risk models for breast cancer since the late 80s. But the way scientists have thought about this has not changed much until recently," says Adam Yala, a doctoral student at MIT CSAIL and co-author of the study, which was published in the medical journal Radiology.

Previous risk models were based on factors including age, family history of breast cancer, breast density, and genetic factors. While these models have helped improve early detection, they miss a lot of important data about the patient and do not provide accurate results at the individual level.

"The problem with this approach is that you're summarizing information that matters before feeding [it] into the model, and this means that the models themselves have not been very accurate," Yala says.

The MIT CSAIL scientists partnered with Massachusetts General Hospital (MGH) and developed a deep-learning model that was trained on 90,000 full-resolution mammogram scans from 60,000 patients who were scanned over the course of several years with various outcomes.

MIT's deep-learning algorithm found patterns in breast tissue that hinted at cancer risk but were too subtle for the human eye to catch. As a result, the AI can find signs of breast cancer in mammograms years earlier than human radiologists, which could reduce invasive treatments and slash medical expenses.

According to the study, the model accurately predicts 31 percent of cancer patients in the highest risk category. The accuracy of existing models stands at around 18 percent.

Providing Personalized Care

One of the benefits of AI-based breast cancer detection is that doctors will be able to provide personalized scanning and prevention for the patients.

"Rather than taking a one-size-fits-all approach, we can personalize screening around a woman's risk of developing cancer," says MIT Professor Regina Barzilay, senior author of the study and a breast cancer survivor herself. "For example, a doctor might recommend that one group of women get a mammogram every other year, while another higher-risk group might get supplemental MRI screening."

"If you give the right screening to the right person, you can both improve the experience and reduce the harms of mammography but also catch the cancers earlier, which makes a huge difference in treatment decisions, because what you do for early stage and late stage cancer are very different," Yala says.

So far, the model has proven to be equally accurate across groups of people of various races and ethnicities. This is one of the pain points of other risk models, whose performance varies across different populations. According to Yala, risk models based on high-level surface factors such as age and family history do not generalize well. For instance, if they're created on data from predominantly white women, they perform poorly on non-white patients.

Recommended by Our Editors

VA Reenlists IBM's Watson in Fight Against Cancer

5 Amazing Projects That Will Change the Future of Healthcare

"Our model is based on the actual patterns in the mammogram. Even though in our data sets, African-American women [comprise] 5 percent of the data set as a whole, the model still performs equally well for both. What that signals to me is that tissue information is more shared, whereas family history might not be," Yala says.

The researchers are now pursuing collaboration with more hospitals to study and serve other groups and make the model even more equitable. They will also be looking to expand the work to other types of cancer, especially those that have less effective risk models, such as pancreatic cancer.

MIT CSAIL's deep-learning model is one of several projects that aim to apply artificial intelligence in the diagnosis and treatment of breast cancer. Large tech companies including IBM, Google, and Alphabet subsidiary DeepMind are leading efforts in the field alongside universities such as New York University and Harvard Medical School.

"Our goal is to make these advancements a part of the standard of care," says Yala. "By predicting who will develop cancer in the future, we can hopefully save lives and catch cancer before symptoms ever arise."

AI tech can identify genetic disorders from a person's face