Mammogram AI Shows Potential in Breast Cancer Screening Research

An AI algorithm may be useful as a second set of electronic eyes for detecting breast cancer, according to a study of mammograms from more than 55,000 women.

Lunit, an organization located in South Korea, created the Insight MMG software, which has received the green light for adoption in the U.S. and Europe. Automatically evaluating breast tissue density, the AI-powered program flags any suspicious nodules.

The Stockholm-based prospective study compared Lunit’s AI to the double-read method, which is recommended by European guidelines and relies on the opinions of two radiologists to arrive at a diagnosis. It looked at the various combinations of specialists and AI that are possible.

In addition to showing that using AI alone does not result in significantly fewer positive breast cancer outcomes, researchers found that using a single human radiologist in tandem with AI software yielded competently similar, if not slightly superior results, with a 4% increase in detection rates.

Researchers found that AI’s single pass could detect nearly as many confirmed instances (98%) as two radiologists, a performance they deemed “non-inferior.” Simultaneously, they discovered that combining AI with the standard two-radiologist approach was more effective. 

Study leader Fredrik Strand, M.D., Ph.D., remarked, “While double readings by two radiologists have been established as the common practice across Europe and Australia, many countries are experiencing great difficulties due to the shortage of radiologists.”

He added that by playing the role of a radiologist, the research sows the seed for widespread usage of AI in breast cancer screening. This can greatly decrease medical costs, he added.

More About the Study

• 11 breast radiologists from Stockholm’s Capio Sankt Göran Hospital participated in the study.
• Their collective experience ranges from 17 to over 30 years. 
• Although radiologists were not privy to the AI’s results before forming their own, the experiment was non-randomized since Lunit’s AI software worked in the background throughout the exams of each patient without altering the hospital’s usual workflow.

Despite AI’s potential to increase the effectiveness of breast cancer screening, its health effects remain unclear, and it’s doubtful that academics will be able to quantify the effect on long-term cancer mortality rates, according to a separate commentary.

Another piece of research using Insight MMG was highlighted by Lunit this week; this one was done in the U.K. and published in the journal Radiology.

This retrospective study instructed the AI on two test sets of mammography data, each having 60 patients’ worth of results that had been assessed by an aggregate of 552 human readers and deemed challenging.

These findings also suggested that there was no discernible difference between the accuracy of the AI software and that of a team consisting of two qualified professionals working together.

Yan Chen, a professor specializing in digital screening at the University of Nottingham, pointed out that, as of now, no other studies have conducted a comparison of human reader performance with AI on such a substantial scale within standard quality assurance test sets. As such, this study could potentially serve as a blueprint for evaluating AI performance in real-world scenarios.