Eisenhower Health

Azadeh Elmi

Clinical AI Modernizer

Dr. Azadeh Elmi is a physician-scientist and breast imaging radiologist dedicated to enhancing patient care through innovation and research. Her work focuses on advancing breast imaging, cancer risk assessment, and personalized screening strategies tailored to individual patient needs. She is particularly interested in leveraging artificial intelligence and emerging imaging technologies to improve diagnostic accuracy, optimize clinical workflows, and shape the future of radiology. Her research and editorial work explore the responsible integration of AI into clinical practice to improve outcomes for patients with breast cancer.

Artificial intelligence (AI) has rapidly transitioned from a conceptual innovation to a noticeable application in daily medical practice. Radiology has emerged as the leading front in AI adoption, accounting for nearly 77 percent of all FDA-approved AI algorithms. Within radiology, breast imaging consistently ranks as the subspecialty most likely to be impacted by AI, reflecting both the important role of pattern recognition for breast cancer detection and high volume of screening examinations.

Diagnostic performance of mammography in the early detection of breast cancer remains a fundamental benchmark in breast imaging. Current data suggest that one of the most promising applications of artificial intelligence in this field is in screening mammography. Large retrospective studies have reported encouraging results, with AI systems claiming to perform at levels approaching radiologists and identifying a significant proportion of early-stage breast cancers. These models primarily leverage advanced patternrecognition techniques, enabling algorithms to detect subtle imaging features and highlight regions suspicious for malignancy. In double-reading environments, replacing one human reader with AI has been associated with reduced recall rates and decreased radiologist workload. Metaanalyses further suggest that radiologists supported by AI outperform standalone readers in both sensitivity and specificity, reinforcing the role of AI as a clinical support tool rather than a replacement.

Despite these encouraging results, real-world implementation has been more gradual and complex than early studies suggested. Surveys from the European Society of Radiology indicate that nearly 70 percent of radiologists report no meaningful decrease in workload following AI adoption, and some note that AI may introduce additional steps rather than streamline daily practice.

Our recent survey of Society of Breast Imaging members similarly highlights a gap between expectations and realworld experience. Only 47 percent of radiologists across the country had already implemented in AI in their clinical practice. Radiologists who have not yet incorporated AI By Azadeh Elmi, MD, Associate Medical Director, Eisenhower Health Azadeh Elmi | | APRIL 2026 9 H UTLOOK ealthcare Tech anticipate meaningful reductions in recall rates, biopsy rates, and burnout. In contrast, radiologists actively using AI report more modest improvements in daily workflow and diagnostic performance after implementation. Based on our survey, while overall attitudes toward AI remain positive, its measurable impact on key clinical metrics appears more variable than initially expected. These findings suggest that benefits demonstrated in controlled studies or vendorreported data may not uniformly translate to routine clinical environments.

“Current AI models, despite strong performance claims, have practical limitations that warrant consideration.”

Current AI models, despite strong performance claims, have practical limitations that warrant consideration. Many commercially available systems rely on either no prior mammograms or at most one to two prior studies, largely because incorporating multiple years of imaging increases computational demands, data storage requirements, and cost. This constrained longitudinal comparison limits the algorithm’s ability to detect subtle interval changes over time—an essential component of early breast cancer detection in routine practice. In addition, performance remains suboptimal in more complex clinical scenarios, particularly in post-surgical breasts and in patients with implants, where architectural distortion, scar tissue, and altered anatomy present challenges those current algorithms do not consistently handle well.

These findings suggest that, although AI is beginning to influence radiology workflows, its impact in breast imaging remains in an early stage. Current AI applications are largely focused on image interpretation, yet many operational inefficiencies occur beyond the reading room. Greater gains in efficiency may come from AI solutions that support patient scheduling, manage callback and follow-up processes, facilitate audit and quality reporting, and improve communication of results and next-step recommendations to both patients and referring physicians.

Importantly, radiologists across experience levels continue to view AI most favorably as a second reader rather than an autonomous decision-maker. This preference reflects the high clinical and emotional stakes of breast imaging, where recall decisions, communication delays, and downstream testing directly affect patient anxiety and outcomes. Workflow-centered AI design, rather than algorithmic performance alone, will be critical to building trust and achieving meaningful adoption.

Looking ahead, the success of AI in breast imaging will depend on significant improvement of the current AI algorithms, thoughtful implementation, continued validation in real-world practice, and closer alignment with radiologists’ operational needs. As AI tools mature, their value will be measured not only by detection metrics but also by their ability to enhance efficiency, consistency, and the overall patient care journey.