AI outperforms humans, spots breast cancer in mammograms more accurately than doctors

The system produced 5.7% fewer false alarms and 9.4% fewer misidentifications; Such a tool could make breast cancer screening more effective, say experts.

                            AI outperforms humans, spots breast cancer in mammograms more accurately than doctors
(Getty Images)

Google's artificial intelligence (AI) program can do a better job than doctors in identifying breast cancer from X-ray images, claims a new study. 

The AI emerged superior by detecting cancers that the radiologists missed in the images and ignoring features that looked like tumors, but were not — the two common errors that creep in when doctors read X-ray images or mammograms, making accurate breast cancer diagnosis a challenge.

If Google's AI software proves to be as successful during clinical trials, it could make breast cancer screening more effective, say experts.

"We hope someday this tool for radiologists becomes as ubiquitous as spell-check for writing e-mail," says Scott McKinney, a Google software engineer and the co-lead author of the study, in a statement.

"This is promising early research, which suggests that in the future it may be possible to make screening more accurate and efficient, which means less waiting and worrying for patients, and better outcomes," Sara Hiom, director of cancer intelligence and early diagnosis at CRUK, told BBC.

Breast cancer is the second leading cause of death from cancer in women. In the US alone, 245,000 women are diagnosed every year, and about 41,000 women die from the disease, according to the Centers for Disease Control and Prevention (CDC). Breast cancer is also the second most common cause of death among white, black, Asian or Pacific Islander, and American Indian/Alaska Native women.

Thousands of mammograms were used to train AI (OPTIMAM)

This could change if breast cancer is caught in its early stages of development. But detecting them early remains a challenge, even for experts. Thousands of cases are missed in scans, including 30% of cancers that develop between screenings. Not to mention, false alarms could arise from mammograms.

False alarms can lead to patient anxiety, unnecessary follow-up and invasive diagnostic procedures. Cancers that are missed at screening may not be identified until they are more advanced — a stage where it cannot be treated, says the study published in Nature.

In order to come up with a better tool, clinicians and radiologists from the Cancer Research UK Imperial Centre joined forces with the AI health research team at Google to develop an AI system.

The first step in this process was to teach the AI system to accurately spot breast cancer in mammograms. So the team fed X-ray images collected from more than 76,000 women in the UK and more than 15,000 women in the US.

After the training, the team tested the AI model by providing the software with X-ray images of women — 25,000 women in the UK and over 3,000 women in the US — who have had either biopsy-proven breast cancer or normal follow-up imaging. The software was also tested against six radiologists.

The results show that AI system outperformed doctors. The trial reduced the number of false alarms from 5.7% to 1.2%. Further, the AI system is as good as two doctors working together, say researchers. This means that this screening tool could either assist or replace a second radiologist. In the UK, two radiologists review mammograms are reviewed by two radiologists, and sometimes a third in case of disagreement.

But, in the US, a single radiologist examines mammograms, leading to more errors. The AI produced 5.7% fewer false alarms and 9.4% fewer misidentifications than the US system.

The study has a few limitations. The study tested only one mammography technology from a single manufacturer, wrote Etta D Pisano from the American College of Radiology, Philadelphia in Nature.

He added that the performance of AI algorithms are highly dependent on the population used in its training. But if AI is tested on a different population, they could make errors. It is therefore important that AI is tailored to the people in a region it is used, he added.

The team hopes that with continued research and clinical studies, the screening tool could increase the accuracy and efficiency of screening programs.

If you have a news scoop or an interesting story for us, please reach out at (323) 421-7514