The integration of artificial intelligence (AI) into mammography screening offers a secure alternative to the traditional approach of radiologists conducting double readings. Furthermore, it has the potential to alleviate the significant burden on medical professionals. These findings were unveiled through a preliminary examination of a prospective, randomized controlled study. This investigation focused on the safety aspect of AI implementation in mammography screening and was overseen by a team of experts from Lund University in Sweden. The comprehensive results of this trial are now available in The Lancet Oncology, marking a substantial advancement in the field.
Annually, nearly one million women in Sweden undergo mammography screening, where each examination is reviewed by two breast radiologists in a double reading process to enhance sensitivity. However, a global shortage of breast radiologists poses a risk to this screening service. Recently, the potential of AI to assist in mammography screening has garnered attention, though its optimal application and clinical impact remain uncertain.
To comprehensively understand the implications of AI-assisted radiology, randomized studies are imperative, assigning women to either AI-supported screening or standard screening. The Mammography Screening with Artificial Intelligence (MASAI) trial is the pioneering randomized controlled study investigating the effects of AI-supported screening.
In this trial, AI was employed to identify high-risk breast cancer screenings for double reading by radiologists. Lower-risk screenings were examined by a single radiologist, with AI highlighting suspicious findings. The trial involved 80,033 women, randomly assigned to either AI-supported screening (40,003) or standard double reading (40,030).
The results indicated that AI increased cancer detection by 20% (41 cases) compared to standard screening, without raising false positive rates. Furthermore, radiologists’ workload was reduced by 44%, with 46,345 AI-supported screenings compared to 83,231 standard screenings.
Of note is the time efficiency, as the study estimated a five-month reduction in radiologists’ time required to review the 40,000 screenings within the AI group.
The study’s context was a single-site Swedish setting, necessitating the validation of these findings under diverse conditions and algorithms.
The next phase involves analyzing the types of cancers detected with and without AI support. The trial’s primary objective is assessing the interval-cancer rate, referring to cancers diagnosed between screenings that typically have poorer prognoses than screen-detected cancers. This assessment will occur following a two-year follow-up for the 100,000 trial participants.
The complexity of screening demands careful consideration of benefits and harms. While higher cancer detection rates do not necessarily equate to a superior method, AI-supported screening appears secure, maintaining cancer detection rates despite a reduced workload. The analysis of interval cancers will unveil whether AI-supported screening enhances accuracy and effectiveness, achieving a balance between overdiagnosis and timely cancer detection.
Find the complete article following the link to The Lancet Oncology.