Scientists train an AI model to predict breast cancer risk from MRI scans

A biopsy that turns out to have benign results can be a relief. But in some cases, it could also mean a patient whose risk of cancer was low from the start has gone through an unnecessarily invasive procedure.

By and large, radiologists recommend that patients whose breast MRI scans raise suspicion of a cancerous growth get a biopsy done. But MRIs often pick up on benign lesions that other mammograms and ultrasound may not. This leads to some patients having their lesions falsely classified as higher risk than they are, and undergoing a biopsy.

In these cases, “radiologists don’t have enough certainty to make a very well-informed decision, should this patient actually undergo a biopsy or not. So they err on the side of, let’s just biopsy a lot of people,” said Jan Witowski, a postdoctoral research fellow at New York University Langone Health. “This is where the hope of using AI tools is.”


In a new paper published recently in Science Translational Medicine, Witowski and his colleagues at NYU and Jagiellonian University in Poland present an artificial intelligence tool that can predict the probability of breast cancer in MRI scans as well as a panel of board-certified radiologists. In a retrospective analysis, it was also capable of reducing unnecessary biopsies by up to 20% for patients whose MRIs show suspicious lesions that might warrant a biopsy, officially known as BI-RADS category 4 lesions.

Outside experts commended the study for developing a promising AI tool to reduce the false-positive rate of breast MRI, noting the rarity of commercial AI tools specifically for processing MRI scans. Reducing the number of benign biopsies would greatly “reduce patient anxiety, subject patients to less unnecessary exams and [procedures], and decrease associated health care costs,” said Manisha Bahl, a breast imaging radiologist at Massachusetts General Hospital and associate professor of radiology at Harvard Medical School.


MRI as a medical imaging technique is the most informationally dense compared to other scans like mammograms and ultrasound, making it the most difficult to process from a technical perspective. Other strengths of the study are the level of detail in the author’s analyses and how the tool can work on external populations.

The system was trained on a dataset of over 20,000 labeled breast MRI scans from one of the NYU Langone Health breast imaging sites, learning to associate certain features in the 3D images with characteristics of breast cancer. It was then externally validated with three datasets, two from the U.S. including one from Duke University, and the third from a site in Poland.

The authors also compared the final performance of the AI system to five board-certified radiologists who regularly read breast MRIs. They found no statistically significant difference in their abilities to discern breast cancer from the images. Following this test, they analyzed whether the AI system can help radiologists to avoid unnecessary biopsies by serving as a guide as they decide the best course of action for a patient.

The model was able to generalize across subgroups stratified by race, age, and different histological and molecular cancer subtypes. “Even in really small subpopulations, this model performed just as well,” Witowski noted. “There were no obvious trends that would suggest that the model doesn’t perform well on even smaller subgroups.”

Adam Yala, an assistant professor of computational precision health at UC Berkeley and UC San Francisco, noted, however, that the training datasets for the model “capture a slice of the overall community of people who use MRI.”

An observation shared by both Yala and Bahl is that MRI images from Black women patients made up only 6% of the NYU training data set, compared to 69% from white women. “Black women were not well-represented in [this] training set, so we wouldn’t necessarily expect that the model performs well in all racial groups, since it’s disproportionally trained based on one group,” Bahl said. Even fewer Asian women were represented, with their images making up 4% of the dataset. White women also accounted for a majority of the patients in the validation datasets.

Yala stressed the importance of having “a much broader, diverse external validation of how this tool performs in very different health centers all over the world.” “If successful, we want to be able to say any hospital that’s interested in using these kinds of tools to help lower their biopsy rates, or better leverage MRI, should be able to use this tool,” he said.

Such a tool will likely aid radiologists’ decision-making about whether to biopsy. Though the standalone system performs just as well as radiologists, it’s difficult to predict how it will work in true clinical settings. The researchers also examined what it might look like, in theory, to put such a model into practice. They did so using a relatively new methodology that explores the risks and benefits compared to potential alternate routes: What if doctors could downgrade some patients with category 4 lesions to category 3, opting them out of a biopsy?

For instance, “they could reduce the number of benign biopsies by a certain amount and not miss any cancers,” Bahl said. “Or they could reduce the number of benign biopsies by a larger amount, but then miss a couple of breast cancers while using that particular threshold.”

To test its true impact in practice, Witowski said, “clinical trials are the next natural step.”

For every breast biopsy that reveals the presence of cancerous cells, two to four other biopsies have been performed and shown benign results. While MRI has traditionally been reserved for high-risk patients, more recent research suggests the technique may be beneficial in women of average or intermediate risk. Bahl noted that as MRI becomes more widely accessible, it will be important to determine an appropriate false-positive rate so that fewer women are subjected to unnecessary biopsies.

“The study is a promising step toward us realizing the potential of MRI and the rich biological information that’s contained within MRI exams,” Bahl said. “Much of that information may be imperceptible to the human radiologist reader, and deep learning offers us the ability to uncover that biological information that’s not discernible to humans.”

Source link

Back to top button