Integration of artificial intelligence (AI) into screening workflows increased the detection of breast cancer by 10.4% in the United Kingdom, according to the results of the GEMINI study published in Nature Cancer. Additionally, use of AI in different workflows led to reductions in workload by up to 31% and savings of 36% compared with standard processes.
Background
In the United Kingdom, all women between the ages of 50 and 70 years are invited for mammograms every 3 years to test for breast cancer, resulting in over 2 million mammograms performed per year.
“Currently, in the [United Kingdom], to reduce the number of cancers missed, two radiologists read every mammogram. However, some breast cancers are extremely hard to detect, and it is not always clear from mammograms whether breast cancer is present. So, when there is the suspicion of cancer on a mammogram the woman is recalled for additional investigations. Despite this, approximately 20% of cancers are missed using this process. Furthermore, many more women are recalled for further assessments than are diagnosed with cancer. For each five women recalled, approximately one will be diagnosed with breast cancer. So, they have had unnecessary, often invasive tests—not to mention the additional worry for the patient,” said lead author Clarisse Florence de Vries, PhD, MSc, Glasgow Lab for Data Science & AI, Public Health, School of Health and Wellbeing, University of Glasgow. “This is why our findings are so important—not only did we find optimal ways to detect breast cancer, quicker, and more accurately, we also found ways to reduce the number of women having to return for unnecessary tests.”
Study Methods
In the GEMINI study, researchers conducted prospective evaluations of 10,889 women within a region of the United Kingdom to determine ways that AI could be used to improve breast cancer screening. Live AI with Mammography Intelligent Assessment (Mia) v.3 was integrated into screening, and the researchers also conducted various simulations to test other ways that AI could be implemented to improve workflows.
All participating women received routine care and their scans were assessed by both the AI tool and a human reader. When the AI tool and human assessment disagreed, the case went on for additional human review.
Key Findings
Disagreement between the AI tool and routine double reading led to the detection of 11 additional cancers.
The resulting cancer detection improvement rate was 10.4%, or 1 per 1,000 patients. The recall rate was reduced by 0.8% and workloads were reduced by up to 31%. Fewer patients were also recalled for biopsies due to false-positive results, thereby further reducing health-care costs and resources as well as patient stress.
Additional variations in the workload to incorporate AI more led to up to 36% in workload savings as well as the superiority of the cancer detection rate, recall rate, positive predictive value, sensitivity, and specificity.
Further, women with detected cancer were able to be notified in a shorter time in the study. The notification time was reduced from 14 days to 3 days, which the study authors considered “hugely significant given that the earlier detection of primarily high-grade cancers enables earlier treatment, which has a greater likelihood of treatment success.”
“Health care and radiology are facing substantial challenges due to high workload, a shortfall of clinical radiologists, and an aging population. However, despite the promise of AI, the UK National Screening Committee does not recommend the use of AI in the NHS breast screening program. They previously highlighted that both the quality and quantity of the evidence base were insufficient,” Dr. de Vries added. “Our work adds high-quality evidence to the scientific literature in support of AI. It also demonstrates that AI use can be tailored to local health-care needs to enhance service delivery.”
Going forward, researchers are expanding this work in the upcoming EDITH trial to evaluate AI use in breast cancer screening across the United Kingdom.
DISCLOSURES: The study was funded through the U.K. National Health Service (NHS) AI in Health and Care Award in partnership with the National Institute for Health and Care Research (NIHR). For full disclosures of the study authors, visit nature.com.
