The study concludes that a combination of Artificial Intelligence algorithms and the interpretations of radiologists could make it so that, in the USA alone, half a million women do not have to undergo unnecessary diagnostic tests every year.
Artificial Intelligence techniques, used in combination with the assessment of expert radiologists, improve the accuracy of detecting cancer with mammograms. This is one of the main conclusions of an international study, conducted by researchers from the Polytechnic University of Valencia (UPV), the Spanish National Research Council (CSIC) and Valencia University (UV), and which has been published in one of the medical journals with the most worldwide dissemination, the Journal of the American Medical Association.
The study is based on the results obtained in the Digital Mammography (DM) DREAM Challenge, an international competition directed by IBM with the participation of the Institute of Corpuscular Physics (IFIC, CSIC-UV) along with scientists of the Institute of Telecommunications and Multimedia Applications (iTEAM) of the UPV.
The team of researchers from the IFIC and iTEAM UPV was the only Spanish group that made it to the final stage of the challenge. To do so, they developed a prediction algorithm from scratch based on convolutional neural networks, an Artificial Intelligence technique that simulates the neurons of the visual cortex and makes it possible to classify images, as well as allowing the system to self-learn. They also used principles linked to the interpretation of X-rays, where the group has several patents. The results of the Valencian team along with the other finalists are the ones published in the Journal of the American Medical Association (JAMA Network Open).
“Having taken part in this challenge has allowed our group to collaborate in Artificial Intelligence projects with clinical groups of the Valencian Community,” highlights Alberto Albiol, professor at the UPV and member of the iTEAM group. “This has opened opportunities for us to apply Machine Learning techniques, as is suggested in the article,” he adds.
For example, the work conducted by the Valencian researchers is being conducted on Artemisa, the new computer platform for Artificial Intelligence of the Institute of Corpuscular Physics financed by the European Union and the Valencia Community Government as part of the FEDER 2014-2020 operative programme for the acquirement of R+D+innovation infrastructure and equipment.
“Designing strategies to decrease the operational costs of the healthcare system is one of the objectives of using Artificial Intelligence sustainably,” highlights Francisco Albiol, researcher of the IFIC who took part in the study. “The challenges go from the algorithmic side to jointly designing strategies based on evidence together with the medical sector. Artificial Intelligence applied on a large scale is one of the most promising technologies to make healthcare sustainable,” he says.
The goal of the Digital Mammography (DM) DREAM Challenge is to involve a broad international scientific community (over 1,200 scientists from around the globe) to assess whether Artificial Intelligence algorithms can match or improve the interpretations of mammograms made by radiologists.
“This DREAM Challenge allowed for the rigorous and appropriate assessment of tens of advanced deep learning algorithms in two independent databases,” explains Justin Guinney, Vice-President of Computational Oncology at Sage Bionetworks and president of the DREAM Challenges.
Half a million less mammograms a year in the USA
Directed by IBM Research, Sage Bionetworks and the Kaiser Permanente Washington Research Institute, the Digital Mammography DREAM Challenge concluded that, despite no algorithm surpassing radiologists on its own, a combination of methods added to the assessments made by experts improved the accuracy of the examinations. Kaiser Permanente Washington (KPW) and the Karolinska Institute (KI) of Sweden provided hundreds of thousands of mammograms and unidentified clinical data.
“Our study suggests that a combination of Artificial Intelligence algorithms and the interpretations of radiologists can make it so that half a million less women every year do not have to undergo unnecessary diagnostic tests in the USA alone,” summarises Gustavo Stolovitzky, director of IBM’s program aimed at Translational Systems Biology and Nanobiotechnology at the Thomas J. Watson Research Centre and founder of the DREAM Challenges.
To ensure the privacy of the data and prevent participants from downloading mammograms with sensitive data, the study organisers applied a work system from the model to the data. In them, participants sent their algorithms to the organisers, who developed a system that applied them directly to the data.
“This approach to share data is particularly innovative and essential to preserve the privacy of the data,” says Diana Buist, from the Kaiser Permanente Washington Health Research Institute. “Furthermore, the inclusion of data from different countries, with different practices when conducting mammograms, reflects important translational differences in the way that Artificial Intelligence could be used for different populations.”
Mammograms are the most commonly used diagnostic technique for the early detection of breast cancer. Although this detection tool is generally effective, mammograms must be assessed and interpreted by a radiologist, who use their human visual perception to identify signs of cancer. Thus, a 10% of ‘false positives’ is estimated among the 40 million women who undergo mammograms every year in the United States.
“An effective Artificial Intelligence algorithm that can raise the ability of the radiologist to decrease the repetition of unnecessary tests while detecting clinically significant cancers would help increase the detective value of the mammograms, efficiently improving the harm-benefit ratio,” believes doctor Christoph Lee, from the Washington School of Medicine.
“Evaluation of Combined Artificial Intelligence and Radiologist Assessment to Interpret Screening Mamograms”, JAMA Network Open. 2020;3(3):e200265. DOI:10.1001/jamanetworkopen.2020.0265