Paris (AFP)

A computer program has shown better accuracy than radiologist experts in identifying breast cancer from mammography images, according to a British study.

Breast cancer is one of the most common cancers in women, with more than two million new cases diagnosed last year worldwide.

These results, published in the scientific journal Nature, "suggest that we are developing a tool that can help doctors pinpoint breast cancer with greater accuracy," notes Dr. Dominic King, UK manager at Google Health, and co-author of this study.

"More trials, clinical validation and regulatory approvals are needed before it can start to make a difference for patients, but we are determined to work with our partners to achieve this goal," added the researcher in a statement. the Imperial College of London.

This artificial intelligence (AI) technique from Google research is based on a mathematical model, an algorithm. The latter was trained, fed, with nearly 29,000 mammography images from Great Britain and to a lesser extent from the United States.

Experts had access to the patient's history when interpreting the radiographic images, while the AI ​​only had access to the last mammogram.

AI has shown a reduction in the proportion of cases where cancer has been wrongly detected, by 5.7% on the American images studied and by 1.2% on the British.

The algorithm also reduced the percentage of missed diagnoses by 9.4% among American images and 2.7% among those from Great Britain.

"The earlier breast cancer is identified, the better it is for the patient," Dominic King, UK head of Google Health, told AFP.

In the United States, only one reading of screening images is generally performed, while in the United Kingdom, mammograms offered to women between 50 and 71 years of age are examined by two radiologists. This is also the case in the context of organized screening offered in France to women aged 50 to 74.

The Google Health team also conducted experiments comparing the computer's decision with that of the first reader radiologist.

If the two diagnoses matched, the case was marked as resolved. It was only in the event of discordant results that the device was then asked to compare with the decision of the second reader.

The study by King and his colleagues shows that using AI to check the diagnosis of the first human reader could save up to 88% of the workload of the second radiologist.

"This technology represents an opportunity to support the excellent work being done by examiners today," said King.

The team hopes that this technology may one day serve as a "second opinion" for cancer diagnoses.

© 2020 AFP