Royal Surrey authored research paper among most popular articles 2020 | News

A research paper showing that artificial intelligence can be more accurate than radiologists in diagnosing breast cancer from mammograms is among the top read articles published in the journal Nature this year.

An international group of researchers, including co-authors Professor Kenneth Young and Professor Mark Halling-Brown from Royal Surrey, designed, trained and evaluated an artificial intelligence (AI) computer model using mammograms from nearly 29,000 women. 

The AI computer program outperformed six radiologists by spotting cancers that they missed, while flagging fewer false positives, which happen when the mammogram is incorrectly diagnosed as abnormal. The AI program was as accurate at reading mammograms as two radiologists working together and was actually superior at spotting cancer than a single radiologist.

In the AI study, compared to one radiologist, there was a reduction of 1.2 per cent in false positives and a reduction of 2.7 per cent in false negatives.

Altmetric, a website that measures attention, influence and impact of publications, ranks this publication as third most popular out of 880 papers published in Nature at a similar time. The paper’s popularity is ranked as 610 out of more than 16million research outputs ever scored by Altmetric – that’s in the top one per cent of all research outputs ever tracked by the system.

The publication has also been mentioned by 177 news outlets including the Guardian and the BBC, 28 blogs and more than 3,000 tweeters.

The study relied on the breast cancer screening Image database, OPTIMAM, which Professor Young first established working with Cancer Research UK more than 10 years ago and that is still growing today. Professor Young said:

“We developed the OPTIMAM database to give researchers and clinicians across the world access to well-documented medical images for the purposes of medical research. Without large well-curated databases like OPTIMAM, it would not have been possible to train successful and generalisable AI algorithms.

“Large amounts of time and funding were needed to develop the automated procedures we use to collect and bring together more than three million images and associated clinical information from several breast screening centres.”

Professor Halling-Brown added:

“The work to collect the de-identified data used in this AI breast cancer research project began more than 10 years ago. The image database is now driving other research and the methodology behind it, which enables clinical data to be collected and de-identified, is being used elsewhere in UK research including a national project studying radiological images of patients with Covid-19.”

Read the publication International evaluation of an AI system for breast cancer screening on Nature’s website.