Easing Workload Pressures While Maintaining Cancer Detection Accuracy
Replacing one human reader in double-reading workflows or using AI as a triage tool could represent a significant step forward in modernizing breast cancer detection
A study out of Denmark evaluating the use of AI in mammography screening demonstrated AI’s ability to significantly reduce the workload of radiologists without diminishing cancer detection accuracy.
As development and adoption of AI tools in the clinical setting advance, performance evaluation is a critical step in ensuring such tools deliver their anticipated benefits. It is also vital for medical imaging professionals to understand the importance of careful integration planning.
“Our work highlights the great potential for AI in making mammography screening more efficient—mainly in terms of workload reduction—and emphasizes the significance of the integration of AI in the screening setting,” said Mohammad T. Elhakim, MD, PhD, a postdoctoral researcher in the Department of Radiology and Research & Innovation Unit of Radiology (UNIFY) at Odense University Hospital.
The research was conducted on a retrospective cohort including more than 249,000 screening mammograms in the Region of Southern Denmark and explored three AI integration scenarios. Two radiologists independently assessed the mammograms, resolving any discrepancies through arbitration.
The findings suggest that AI could replace one or both readers in specific roles:
- Replacing the first reader: AI fully replaced the first radiologist, maintaining cancer detection rates while reducing the volume of screening reads by 48.8%.
- Replacing the second reader: AI replaced the second radiologist, cutting workload by 48.7% and lowering recall rates by 2.2%. However, the researchers note, this scenario showed a slight reduction in sensitivity (−1.5%, p<.001).
- AI as a triage tool: AI independently assessed low-risk (no recall) and high-risk (recall) cases, referring only moderate-risk cases for human review. This approach achieved the greatest workload reduction (49.7%) and slightly improved cancer detection rates compared to standard methods.
“This study represents one of the most extensive investigations to date into the use of AI in a real-world screening context,” said Abhinav Suri, MPH, a medical student at the David Geffen School of Medicine at the University of California, Los Angeles, and author of a related commentary in Radiology: Artificial Intelligence. “Most notably, the number of mammograms requiring human review was almost halved across all three scenarios, demonstrating that AI can effectively prioritize cases without compromising diagnostic accuracy,” Suri said.
The practice of performing double readings on mammograms is well established in many countries across the globe where it has been shown to yield improved detection rates. It is not standard practice in the United States.
According to Suri, who earned undergraduate degrees in computer science and biology from the University of Pennsylvania in Philadelphia, the advent of AI promised to alleviate challenges faced by radiologists by serving as an additional “pair of eyes” or even by replacing one of the human readers in the double-reading process.“As AI systems begin to be deployed in clinical environments, there is an urgent need to rigorously evaluate their performance using large-scale, representative datasets.”
— ABHINAV SURI, MPH
These roles for AI raise several important questions and necessitate closer scrutiny.
“As AI systems begin to be deployed in clinical environments, there is an urgent need to rigorously evaluate their performance using large-scale, representative datasets,” he said.
Findings from the study answer that need and underscore the potential of AI to streamline breast cancer screening processes. They also highlight its sometimes-surprising ability to perform as well as, or better than, human radiologists under specific conditions.
“We did not expect the triage-based AI integrated scenario to have significantly higher detection rates in comparison to the other scenarios, while still reducing workload,” Dr. Elhakim said. “Yet, replacing the second reader resulted in minor trade-offs, such as decreased sensitivity, which could impact the detection of certain cancers.”
Based on the outcome of the study, Dr. Elhakim urged those considering AI solutions to determine the optimal integration point for AI in screening workflows to best capture its benefits without degrading diagnostic accuracy.
Considerations for Real-World Application
While the results are promising, Dr. Elhakim and colleagues acknowledge the study’s limitations.
“These scenarios were simulated using retrospective data, and real-world implementation may reveal additional challenges, including potential automation bias,” they noted.
Suri agreed, emphasizing important contextual considerations to consider when evaluating the study outcome. “The results in this study only pertain to one AI software product, which will limit the applicability of these results to other vendors,” he said. “Additionally, the population on which the AI was simulated was a cohort of patients in southern Denmark, which may not apply to populations in other regions.”
Dr. Elhakim and colleagues also noted that differences in radiologist experience levels, which influenced outcomes in the study, could shift with AI adoption. And, while AI holds great promise, legal, ethical and practical considerations must be addressed before widespread implementation.
“For instance, national guidelines currently prohibit the full replacement of human radiologists, making a hybrid approach—such as replacing the first reader—a more feasible solution,” Dr. Elhakim said.
With this research, Dr. Elhakim and his team add to a growing body of evidence supporting AI’s role in medical imaging. AI could help address the increasing demand for screening services and mitigate radiologist shortages.
Of course, optimal patient care remains a central concern. “As AI technology continues to advance and become more integrated into clinical workflows, ongoing research and rigorous evaluation will be essential to ensure that these systems are used safely and effectively,” Suri said.
“Based on the findings of the current study, a process for AI implementation in clinical screening practice is ongoing, and on this basis, more research will potentially be conducted for post-implementation monitoring,” Dr. Elhakim said.
The findings pave the way for new prospective trials to validate these scenarios in real-life settings and add to the findings of prospective trials already underway or completed. The researchers stress that future studies will need to explore long-term outcomes, such as the impact on interval cancers, and assess how AI integration influences clinical workflows and radiologist decision-making.
“Prospective validation studies are crucial in translating these simulated scenarios into practical, real-world solutions,” Dr. Elhakim said.
For More Information
Access the Radiology: Artificial Intelligence article, “Artificial Intelligence: AI-integrated Screening to Replace Double Reading of Mammograms: A Population-wide Accuracy and Feasibility Study,” and the related commentary, “AI as a Second Reader Can Reduce Radiologists’ Workload and Increase Accuracy in Screening Mammography.”
Read previous RSNA News stories on breast imaging: