New transparent AI predicts breast cancer in 5 years

New ways to use artificial intelligence (AI) make predictions possible. breast cancer Predicted 5 years in advance with amazing accuracy. And unlike previous AI models, we know how this model works.

A new AI system called AsymMirai simplifies previous models by only comparing the differences between the left and right breasts to predict risk. Developers say it has the potential to save lives, prevent unnecessary testing and save money for the healthcare system.

“With traditional AI, you ask a question and it spits out an answer, but no one actually knows how the AI ​​makes decisions. It's a black box,” says John, a doctoral student in Duke University's Department of Computer Science. Donnelly says: , Durham, North Carolina, first author. new paper in Radiology Model description.

“Our approach allows people to see how the algorithm produces the output, so they can fact-check and trust it,” he said.

One in eight women will develop invasive breast cancer, and one in 39 will die from invasive breast cancer.mammography misses approx. 20% About breast cancer. (Actress Olivia Munn recently highlighted the shortcomings of genetic screening and mammography. disclosed Although she was in normal condition, she was undergoing treatment for aggressive breast cancer. mammogram Genetic testing is negative. )

Vivian Freitas, MD, a radiologist and assistant professor of medical imaging at the University of Toronto in Canada, said the model could help doctors bring the abstract idea of ​​AI to the bedside in a meaningful way. .

“This is a new chapter in the field of AI,” Freitas said. Editor Praise the new paper. “This makes AI more tangible, easier to understand, and more likely to be accepted.”

AI as a second eye

Mr. Donnelly described AsymMirai as a simpler, more transparent, and easier-to-use version of Mirai. Mirai is a breakthrough AI model that made headlines in 2021 with its promise to determine with unprecedented accuracy whether a patient is likely to develop breast cancer within the next five years. Year.

Mirai has been identified Up to double the number of future cancer diagnoses compared to the traditional risk calculator Tyrer-Cuzick. Accuracy was also maintained across diverse patient populations. This is a notable plus for two fields that have notoriously poor outcomes for minorities: AI and healthcare.

Tyrer-Cuzick and other low-tech risk calculators use your personal and family history to statistically calculate your risk. Mirai, on the other hand, analyzes the millions of raw data embedded in mammograms to decipher patterns that are invisible to radiologists. Four images of her, including two angles from each breast, are fed into the model, which generates a score from 0 to 1 indicating her risk of developing breast cancer within 1, 3, or 5 years.

but Mirai creators too He admits he didn't know exactly how that score was arrived at, a fact that adds to the hesitation among clinicians.

Study co-author Fides Schwartz, M.D., a radiologist at Brigham and Women's Hospital in Boston, Massachusetts, said researchers were able to crack the code on Mirai's “black box” and that the scores were largely positive. He said the findings were determined by evaluating subtle differences between the breast tissue and the right breast tissue. Left breast tissue.

Knowing this, the research team simplified their model to predict risk based only on “local two-way dissimilarities.” Asim Mirai was born.

The team then used AsymMirai to look back at more than 200,000 mammograms from approximately 82,000 patients. They found that it worked much like its predecessor, assigning people a higher risk of developing cancer 66% of the time (compared to 71% for Mirai). The effect was even better for patients who noticed the same asymmetry for multiple years in a row: He was 88% likely to give a higher score to those who later developed cancer than those who did not.

“We found that we could predict with surprisingly high accuracy whether a woman would develop cancer within the next one to five years, based solely on local differences between left and right breast tissue,” Donnelly said. Masu.

Schwartz envisions a day when radiologists can use this model to develop personalized screening strategies for their patients. Doctors may advise patients with high scores to get screened more often than guidelines suggest, add her MRI to a mammogram, and closely monitor problem areas identified by AI. .

“On the other hand, for people who are really low risk, it may be possible to skip annual testing, which may not be very comfortable and may not be necessary,” Schwartz said.

cautious optimism

Dr. Robert Smith, senior vice president of early cancer detection science at the American Cancer Society, said AI has been used for decades to reduce the workload of radiologists and improve diagnosis.

“But AI has not been able to reach its full potential,” Smith said. “Inexperienced radiologists often used AI as a crutch. Instead of reading mammograms and confirming the AI’s opinion, they ended up letting the AI ​​do the work for them. ” Most of the work was done, but frankly, it wasn't all that accurate. ”

She hopes that newer, more sophisticated AI medical imaging platforms (approximately 18 to 20 models are in development) can eventually save women's lives, especially in areas with a shortage of radiologists. I am.

But he thinks it will be a long time before doctors and patients risk postponing mammograms based on algorithms.

“We just haven't gotten there yet,” he said.