Dynamic duo: AI teamed with radiologist diagnoses breast cancer with 90% accuracy

New York University scientists, teamed with artificial intelligence, have been able to accurately diagnose breast cancer at a 90% clip, speeding up treatment and hopefully reducing the need for biopsies.

That’s according to a new study out of NYU’s School of Medicine, published earlier this month in the journal IEEE Transactions on Medical Imaging. Researchers trained the machine by feeding it almost 230,000 digital mammography exams and more than 1 million images. Radiologists and the computer were able to play off of each other’s strengths, with AI pinpointing pixel-level changes unseen by the eye, and humans using reasoning to reach conclusions missed by the system.

Scientists stressed that their goal is to augment the work of imaging professionals, not replace them. They’re moving slowly to adopt machine learning in clinical practice as the machine continues to build knowledge.

“The transition to AI support in diagnostic radiology should proceed like the adoption of self-driving cars—slowly and carefully, building trust and improving systems along the way with a focus on safety,” first author Nan Wu, a doctoral candidate at the NYU Center for Data Science, said in a statement.

The study’s authors noted that in 2014 alone, U.S. providers performed more than 39 million mammography exams to screen those without cancer symptoms. Doctors referred women with abnormal results for a biopsy to remove a small sample of breast tissue, a procedure that can be painful, costly and increase anxiety for patients.

NYU said its goal is to eventually limit the number of unnecessary biopsies by using AI to speed up diagnosis. They hope to do so by increasing physicians’ confidence in the accuracy of assessments made for screening exams.

Researchers trained the system to analyze images from a database of patients for which a cancer diagnosis had already been reached, testing to see if it could reach the same conclusions. They plan to further improve accuracy by training the program with even more data, and teaching it to spot tissue that is not yet cancerous, but has the potential to reach that stage.